JP5548990B2 - Evolutionary condition judgment network - Google Patents

Description

  The present invention relates to a method of learning a condition judgment network obtained by developing a condition judgment tree by evolutionary computation.

  The inventor has long studied the fully automatic construction / optimization technology for image processing / recognition and has achieved many results. For example, we have proposed a learning method that can automatically construct image processing logic simply by providing an original image and a target image, and a learning method that can automatically construct image recognition logic simply by providing an image and its class. .

  However, the generated processing logic itself has been optimized, and although it has excellent performance, it is often difficult for humans to understand. For this reason, there was a case where vague anxiety was held.

  Therefore, the task was to generate easy-to-understand logic through learning in line with human logical thinking. Examples that are easy to accept as human thoughts include the inevitable three-stage theory and decision trees such as “B for A, C for B, so C for A”. In fact, these methods also contain uncertainties, but they are often acceptable and can be accepted.

Tomoharu Nagao “Evolutionary Image Processing” Shoshodo, May 15, 2002, p. 79-140

  In the present invention, priority is given not only to automatic construction of a high-performance classifier, but also to a human being who understands what has been completed and can be used with consent.

The condition judgment network computing device according to the present invention is:
Target multidimensional data using a condition judgment network that links the start of target multidimensional data to be determined, the start of reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value A condition determination network computing device that determines an evaluation value for target multidimensional data based on data and reference multidimensional data, and has the following elements: (1) a feature amount ID for each intermediate node; Condition determination network individual structure storage unit for storing condition determination network individual structure information having node information capable of specifying a plurality of link destinations and a plurality of threshold values. For each operation node, calculate a feature quantity for the reference multidimensional data according to the feature quantity ID, compare the feature quantity with one of the threshold values, The process of selecting a node to be a link destination based on the result is repeated, and a reference multidimensional data path determination unit that generates a path of reference data (3) specifies operation nodes sequentially from the start of the target multidimensional data along the link, For each computation node, the feature amount for the target multidimensional data is calculated according to the feature amount ID, a threshold is selected based on whether the computation node is a node on the reference data path, and the feature amount is compared with the selected threshold. And a target multidimensional data path determination unit that identifies an evaluation value corresponding to the terminal when the terminal is reached by repeating the process of selecting a link destination node based on the comparison result.

Also,
The node information of the condition determination network individual structure information can specify a threshold value when receiving a standard threshold value and an action as a plurality of threshold values,
The target multidimensional data path determination unit selects a standard threshold value when the computation node is not a node on the reference data path, and sets a threshold value when the computation node is affected when the computation node is a node on the reference data path. It is characterized by selecting.

Also,
The target multidimensional data is image data to be determined, and the reference multidimensional data is image data to be referred to.

The condition determination network generation system according to the present invention is:
A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation A determination network generation system having the following elements: (1) A sample data set storage unit (2) for storing as many sample data sets as target multidimensional data and reference multidimensional data A target evaluation value storage unit that stores a target evaluation value corresponding to each sample data set. (3) A condition determination network individual having node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of thresholds for each intermediate node. Individual group storage unit for storing a group of structure information (4) For each condition determination network individual structure information,
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
The fitness calculation unit for calculating the fitness of the condition determination network individual structure information based on the number of correct answers with respect to the number of samples (5) Based on the fitness, from the group of the condition determination network individual structure stored in the individual group storage unit Individual selection unit for selecting a part (6) Reproductive unit for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (7) Condition determination network individual structure based on fitness A learning individual output unit that selects and outputs.

The condition determination network calculation method according to the present invention is:
In order to define a condition judgment network that links the start of the target multidimensional data to be determined, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value Each node has a condition determination network individual structure storage unit for storing condition determination network individual structure information having node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values, and is referred to the target multidimensional data A condition determination network calculation method by a condition determination network calculation device for determining an evaluation value for target multidimensional data based on multidimensional data, comprising the following elements: (1) Reference multidimensional data start The operation nodes are sequentially identified along the links from the first to the second, and the feature amount for the reference multidimensional data is determined according to the feature amount ID for each operation node. A reference multidimensional data path determination step for generating a reference data path by calculating, comparing the feature quantity with one of the threshold values, and repeatedly selecting a link destination node based on the comparison result. The calculation nodes are sequentially identified along the link from the beginning of the dimensional data, the feature quantity for the target multidimensional data is calculated for each calculation node according to the feature quantity ID, and whether the calculation node is a node on the reference data path or not The threshold value is selected based on the comparison, the feature value is compared with the selected threshold value, and the process of selecting the link destination node based on the comparison result is repeated, and when the termination is reached, the evaluation value corresponding to the termination is specified. Target multidimensional data path determination step.

The program according to the present invention is:
In order to define a condition judgment network that links the start of the target multidimensional data to be determined, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value Each node has a condition determination network individual structure storage unit for storing condition determination network individual structure information having node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values, and is referred to the target multidimensional data A computer that is a condition determination network computing device that determines an evaluation value for target multidimensional data based on multidimensional data is caused to execute the following procedure: (1) Along the link from the beginning of the reference multidimensional data The operation nodes are sequentially identified, and the feature amount for the reference multidimensional data is calculated according to the feature amount ID for each operation node. A reference multidimensional data path determination procedure for generating a reference data path by comparing the threshold values of the shifts and repeating the process of selecting a link destination node based on the comparison result. (2) From the start of the target multidimensional data to the link The calculation nodes are sequentially identified along the line, the feature quantity for the target multidimensional data is calculated for each calculation node according to the feature quantity ID, and a threshold is selected based on whether the calculation node is a node on the reference data path. The target multidimensional data path determination that compares the feature value with the selected threshold and repeats the process of selecting the link destination node based on the comparison result, and specifies the evaluation value corresponding to the terminal when it reaches the terminal procedure.

The condition determination network generation method according to the present invention is:
A sample data set storage unit for storing sample data sets composed of target multidimensional data and reference multidimensional data for the number of samples;
A target evaluation value storage unit for storing a target evaluation value corresponding to each sample data set;
An individual group storage unit that stores a group of condition determination network individual structure information including node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values for each intermediate node;
A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation A condition determination network generation method by a determination network generation system, comprising the following elements: (1) For each condition determination network individual structure information,
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
A fitness calculation step for calculating the fitness of the condition determination network individual structure information based on the number of correct answers with respect to the number of samples (2) Based on the fitness, from the group of the condition determination network individual structure stored in the individual group storage unit Individual selection process for selecting a part (3) Reproductive process for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (4) Condition determination network individual structure based on fitness A learning individual output step of selecting and outputting.

The program according to the present invention is:
A sample data set storage unit for storing sample data sets composed of target multidimensional data and reference multidimensional data for the number of samples;
A target evaluation value storage unit for storing a target evaluation value corresponding to each sample data set;
An individual group storage unit that stores a group of condition determination network individual structure information including node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values for each intermediate node;
A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation (1) For each condition determination network individual structure information, which causes a computer to be a determination network generation system to execute the following procedure:
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
Fitness calculation procedure for calculating fitness of condition determination network individual structure information based on the number of correct answers with respect to the number of samples (2) Based on fitness, from the group of condition determination network individual structures stored in the individual group storage unit Individual selection procedure for selecting a part (3) Reproductive procedure for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (4) Condition determination network individual structure based on fitness Learning individual output procedure to select and output.

  In a normal decision tree, a condition is determined for data to be determined, and a path procedure for determining a destination of branching is taken. In the present invention, a path procedure is also taken for reference data. Then, the threshold value is changed for the nodes on the route. Thereby, the condition judgment in the data to be judged is subjected to the condition judgment by the reference data. Themes that cannot be solved only by absolute judgment on the data to be judged can be solved by taking into account relative judgment due to the influence of the reference data.

  Since the present invention is a procedure that includes and evolves a condition judgment tree (decision tree), it is expected to be easily understood by humans.

FIG. 1 is a diagram illustrating a configuration of a condition determination network generation system. FIG. 2 is a diagram showing a flow of learning processing by the evolutionary calculation algorithm. FIG. 3 is a diagram illustrating an example of a block image. FIG. 4 is a diagram illustrating an example of the condition determination network individual structure. FIG. 5 is a diagram illustrating a configuration of the condition determination network. FIG. 6 is a diagram illustrating a configuration of the fitness calculation unit. FIG. 7 is a diagram showing a fitness calculation processing flow. FIG. 8 is a diagram illustrating a configuration of the individual calculation unit. FIG. 9 is a diagram showing an individual calculation processing flow. FIG. 10 is a diagram illustrating a reference block image path determination processing flow. FIG. 11 is a diagram illustrating an example (part 1) of the reference block image path. FIG. 12 is a diagram illustrating an example (part 1) of the reference block path table. FIG. 13 is a diagram illustrating a target block image path determination processing flow. FIG. 14 is a diagram illustrating an example (part 1) of the target block image path. FIG. 15 is a diagram illustrating an example (part 2) of the reference block image path. FIG. 16 is a diagram illustrating an example (part 2) of the reference block path table. FIG. 17 is a diagram illustrating an example (part 2) of the target block image path. FIG. 18 is a diagram showing a reproductive processing flow. FIG. 19 is a diagram showing a crossover process flow. FIG. 20 is a diagram showing a mutation process flow. FIG. 21 is a diagram showing a life extension process flow. FIG. 22 is a diagram illustrating an example of a condition determination network individual structure according to the second embodiment. FIG. 23 is a diagram illustrating an example of the condition determination network individual structure according to the third embodiment. FIG. 24 is a diagram illustrating an example of the condition determination network individual structure according to the fourth embodiment. FIG. 25 is a diagram illustrating an example of the condition determination network individual structure according to the fifth embodiment. FIG. 26 is a diagram illustrating an example of the condition determination network individual structure according to the sixth embodiment. FIG. 27 is a diagram illustrating a configuration of a condition determination network according to the seventh embodiment. FIG. 28 is a diagram illustrating an example of the condition determination network individual structure according to the seventh embodiment. FIG. 29 is a diagram showing a node specifying process flow based on the standard threshold. FIG. 30 is a diagram showing a target block image path determination processing flow according to the seventh embodiment. FIG. 31 is a diagram showing a node specifying process flow based on the action threshold.

  FIG. 1 is a diagram illustrating a configuration of a condition determination network generation system. The condition determination network generation system includes an initial individual input unit 101, an individual group storage unit 102, an fitness calculation unit 103, a sample data set storage unit 104, a target evaluation value storage unit 105, an fitness storage unit 106, an individual selection unit 107, A selected individual storage unit 108, a reproductive unit 109, and a learning individual output unit 110 are included.

  In this system, a condition judgment network is generated by learning processing using an evolutionary calculation algorithm. FIG. 2 is a diagram showing a flow of learning processing by the evolutionary calculation algorithm.

  An initial individual input process (S201) by the initial individual input unit 101 is performed. In this process, an initial individual group is input and stored in the individual group storage unit 102. In other words, the condition determination network individual structure information (hereinafter referred to as the condition determination network individual structure) for specifying the structure of the network type individual for which the condition determination is performed is input by the number of elements constituting the group, and the individual identification is performed on them. Information (hereinafter, referred to as individual ID) is associated and stored in the individual group storage unit 102.

  The condition determination network specified by the condition determination network individual structure evaluates the target block image using the reference block image as an auxiliary. FIG. 3 is a diagram illustrating an example of a block image. It is a photographic image of the leather surface. It is divided into three stages, upper, middle, and lower, and each stage is divided into left, center, and right blocks. An example will be described below in which an image selected from these nine block images is set in advance as a reference block image and a target block image, and whether the target block image has a flaw on the leather. The reference block image is referred to as data indicating a leather base.

  An example of the condition determination network individual structure will be described. FIG. 4 is a diagram illustrating an example of the condition determination network individual structure. FIG. 5 is a diagram illustrating a configuration of the condition determination network. The condition determination network will be described based on this figure.

  The condition determination network individual structure has node information (records) 451 to 457 for each node. The node information is associated with a node number 401 for identifying a node, a feature amount ID 402 for identifying a feature amount calculated by the node, and a small side link that is a link destination node number when the feature amount is smaller than a threshold value A destination 403, a large link destination 404 that is a node number to which a link is made when a feature amount is larger than a threshold, a standard threshold 405 used as an initial value, and an action threshold 406 used instead of the initial value are stored. .

  In this example, node numbers 1 and 2 are also assigned to the start end. Node number 1 is the reference block image starting point that is the starting point of the determination path for the reference block image, and node number 2 is the starting point of the target block image that is the starting point of the determination path for the target block image. In the case of the start end, since there is one link destination, only the small side link destination 403 is used. The large link 404, the standard threshold 405, and the action threshold 406 are ignored. The large-side link destination 404 may be used.

  In this example, node numbers 6 and 7 are also assigned to the end points. This is the end point of the determination path related to the reference block image and the determination path related to the target block image. In particular, when the end point of the determination path for the target block image is reached, the determination result is assigned to OK or NG depending on the type of OK end and NG end. OK and NG are examples of evaluation values (belonging classes) to be selected. The small side link destination 403, the large side link destination 404, the standard threshold value 405, and the action threshold value 406 are ignored.

  When the link destination of the reference block image start end and the target block image start end link destination are fixed, these start end records can be omitted. Also, the link destination of the reference block image start end and the target block image start end link destination can be made the same. Furthermore, when the node numbers of the OK termination and the NG termination are fixed, the records of these terminations can be omitted.

  This example is expressed with a small number of intermediate nodes for the sake of brevity, but more intermediate nodes are actually set. For example, an intermediate node of about 30 to 40 is set. However, the number of intermediate nodes is not particularly limited. An appropriate number can be selected in accordance with the assumed scale of logic.

  The determination process of the determination path regarding the reference block image and the determination path regarding the target block image will be described later with reference to FIGS.

  Following the initial individual input process (S201) in FIG. 2, the fitness calculation process (S202) by the fitness calculation unit 103 is performed. In this process, the fitness of each individual stored in the individual group storage unit 102 is calculated. The fitness is a value indicating the degree to which the individual is adapted to the target of the sample group. More specifically, a learning sample data set is input, and the ratio at which the judgment evaluation value, which is the result of calculation according to the calculation procedure specified by the individual structure, matches the target evaluation value to be used is used. The sample data set storage unit 104 stores as many sample data sets including the target block image and the reference block image as the number of samples. The target evaluation value storage unit 105 stores a target evaluation value corresponding to each sample data set. The target evaluation value can be not only binary but also three or more values. The calculated fitness is stored in the fitness storage unit 106 in association with the individual ID. Details of the fitness calculation process (S202) will be described later.

  Subsequently, an individual selection process (S203) by the individual selection unit 107 is performed. In this process, an individual to be subjected to reproductive processing is selected by a predetermined number based on the fitness of each individual. For example, a selected number of individuals are selected in descending order of fitness. Alternatively, the number of selections is divided into the number of highly adaptive individuals selected in the high fitness layer and the number of low adaptive individuals selected in the low fitness layer, and the individuals with the high adaptive individual selection number are selected in descending order of fitness. Further, the fitness may be stratified so that individuals with a low adaptive individual selection number are selected in order of lower fitness, and a predetermined number of each may be selected from each layer. In addition, a roulette selection method that selects individuals with a probability proportional to fitness, a rank selection method that determines the number of selections according to the fitness ranking, and a random number of individuals selected as the tournament size from a population. There is also a tournament selection method in which an individual with the highest fitness among them is selected as a tournament winner. In this way, the ID of the selected individual group is stored in the selected individual storage unit 108.

  Subsequently, reproduction processing (S204) by the reproduction unit 109 is performed. In this process, a configuration related to generational change is applied to an individual extracted from the selected individual group. In this example, crossover processing for exchanging some node information (record) between two individuals, mutation processing for changing some items in node information, life extension processing for maintaining the current solid structure To do. Through these processes, a next-generation individual group consisting of individuals of a predetermined number of next-generation individuals is generated, and the condition determination network individual structure is stored in the individual group storage unit 102 in association with each individual ID. Details of this reproductive process (S204) will be described later.

  Then, the generation change process (S202 to S204) is repeated, and when it is determined that the generation change end determination process (S205) by the generation change end determination unit (not shown) continues, the generation change process is looped and ended. If it is determined, the generation change process loop is terminated and the process proceeds to post-processing. The end of the generation change is determined by the number of generations, for example. The number of generation change processes is counted, and when the number of generations is less than the predetermined number of generations, it is determined to be continued, and when the number of generations is reached, it is determined to end. Alternatively, it can be determined that the fitness is complete when the fitness satisfies the termination condition. For example, the target fitness and the target number of individuals are set as end conditions, the number of individuals that have reached the target fitness is counted, and if the number does not reach the target number of individuals, it is determined to continue and the target number of individuals has been reached. It is determined that the process ends. Alternatively, set a time limit for generation change processing, measure the processing time between the loops of generation change processing, determine if the processing time does not reach the time limit, and decide to end when the time limit is reached To do.

  As post-processing, fitness calculation processing (S206) by the fitness calculation unit 103 is performed on the individual population of the last generation as described above, and learning individual output processing (S207) by the learning individual output unit 110 is performed. In the learning individual output process, the condition determination network individual structure of each individual included in the individual population of the final generation and the fitness of the individual are output in association with each other. That is, for each individual ID stored in the individual group storage unit 102, the condition determination network individual structure corresponding to the ID is extracted, and the fitness associated with the ID is extracted from the fitness storage unit 108. The extracted condition determination network individual structure and fitness are associated with each other and output as learning individual information. At that time, it is also effective to sort the individual information group using the fitness as a key and output the sorting result.

  Hereinafter, the fitness calculation processing (S202, S206) and the reproduction processing (S204) shown in FIG. 2 will be described in detail.

  First, the fitness calculation processing (S202, S206) will be described. FIG. 6 is a diagram illustrating a configuration of the fitness calculation unit. The fitness calculation unit 103 includes an individual calculation unit 601, a correct answer determination unit 602, a correct answer number counter 603, and a correct answer rate calculation unit 604.

  FIG. 7 is a diagram showing a fitness calculation processing flow. The processing of S702 to S706 is repeated for each individual stored in the individual group storage unit 102, that is, for each condition determination network structure (S701), and the fitness (correct answer rate) for each individual is obtained.

  As processing for one individual, individual calculation processing (S703) by the individual calculation unit 601 and correct answer determination processing (S704) by the correct answer determination unit 602 are repeated for each sample (S702). As a result, the proportion of correct answers in all samples can be obtained.

  In the individual calculation process (S703), the target block image and the reference block image are input as a sample data set as shown in FIG. 6, and a determination evaluation value is obtained as a determination result by the condition determination network. This will be described later with reference to FIGS.

  In the correct answer determination process (S704) by the correct answer determination unit 602, the determination evaluation value (eg, OK / NG) calculated in the individual calculation process (S703) matches the target evaluation value (eg, OK / NG) of the sample. Determine whether. If they match, the number is counted as a correct answer. Specifically, in the case of a correct answer, the value of the correct answer counter 603 is incremented. The correct answer counter 603 is reset to an initial value 0 for each loop of the condition determination network individual structure. That is, initialization is performed in S701. When all the samples have been processed (S705), the correct answer rate calculation process (S706) by the correct answer rate calculating unit 604 obtains the number of samples from the sample data set storage unit 104, reads the correct answer number from the correct answer number counter 603, and the correct answer number. Is divided by the number of samples to obtain the correct answer rate as a quotient. This correct answer rate is the fitness of the condition determination network individual structure. These processes are completed for all individuals (S707).

  The individual calculation process (S703) described above will be described in detail. FIG. 8 is a diagram illustrating a configuration of the individual calculation unit. The individual calculation unit 601 includes a reference block image path determination unit 801, a reference block path table 802, and a target block image path determination unit 803. In the reference block path table 802, the node that becomes the reference block image path is recorded. Therefore, a reference block path flag is stored for each node in association with the node number (FIG. 12).

  FIG. 9 is a diagram showing an individual calculation processing flow. Processing is performed in the order of reference block image path determination processing (S901) by the reference block image path determination unit 801 and target block image path determination processing (S902) by the target block image path determination unit 803. The reference block path table 802 is initialized for each sample. That is, as a pre-process for starting the reference block image path determination process (S901), it is reset to the initial value.

  First, the reference block image path determination process (S901) will be described. In this process, for the reference block image, a route is selected by feature amount determination at a node of the condition determination network, and the route is recorded. FIG. 10 is a diagram illustrating a reference block image path determination processing flow. First, an operation node is specified from the link destination of the reference block start end (S1001). For example, in the case of the condition determination network individual structure shown in FIG. 4, since node number 1 is the reference block start end, the link destination in the node information (in this example, the small side link destination is used) “ The node number 3 to be calculated next is specified by reading “3”. Then, the reference block route flag corresponding to the operation node in the reference block route table 802 is set to ON (in the route) (S1002). Then, the feature amount ID of the operation node is read (S1003), and the feature amount for the reference block image is calculated (S1004). For example, feature quantities such as the average value, median value, and mode value of the gradation values of the pixels included in the image are calculated. Then, the calculated feature amount is compared with the standard threshold value of the operation node (S1005). For example, if the feature amount is smaller than the standard threshold value, an operation node is identified from the smaller link destination (S1006). In this example, “less than” is a condition, but “below” may be a condition. On the other hand, if the feature amount is larger than the standard threshold value, an operation node is specified from the larger link destination (S1007). In this example, “more than” is a condition, but “greater than” may be a condition. If the identified next computation node is an OK termination or an NG termination (S1008), the reference block path flag at the termination is turned ON and the processing ends. On the other hand, if neither the OK termination nor the NG termination, the processing from S1002 is repeated. Thereby, the path | route with respect to a reference block image is formed.

  FIG. 11 is a diagram illustrating an example (part 1) of the reference block image path. A straight link line indicates a reference block image path. In this example, the route is in the order of node number 1, node number 3, and node number 6.

  FIG. 12 is a diagram illustrating an example (part 1) of the reference block path table. OFF means a node outside the route, and ON means a node inside the route. The initial value is OFF, and all are OFF in the initialization stage. Node number 1 is set to ON (within the path) at the stage of reading the link destination from the reference block start end (S1001 in FIG. 10). The reference block path flag for the node number that is the path among the intermediate nodes is turned ON. Note that the reference block path flags at the start and end may be omitted.

  FIG. 11 and FIG. 12 show an example where the upper central block of the block images shown in FIG. 3 is used as a reference block image. Since the average value of the pixel gradation regarding the upper center block is “146”, the link on the large side is traced at the node number 3 and the path is closed at the OK terminal.

  In the above example, all standard threshold values are used, but it is also possible to operate so as to use all action threshold values. Alternatively, it is also conceivable to add a threshold dedicated to the reference block to the node information and use the dedicated threshold.

  Next, the target block image path determination process (S902) will be described. FIG. 13 is a diagram illustrating a target block image path determination processing flow. First, an operation node is identified from the link destination at the start end of the target block (S1301). For example, in the case of the condition determination network individual structure shown in FIG. 4, since node number 2 is the start of the target block, the link destination in the node information (in this example, the small side link destination is used). The node number 4 to be calculated next is specified by reading “4”. Then, the feature amount ID of the operation node is read (S1302), and the feature amount of the target block image is calculated (S1303). Here, the reference block route flag (S1304) corresponding to the computation node is read from the reference block route table 802, and when it is OFF (outside route), the standard threshold value of the computation node is read (S1305), and the feature amount and the standard threshold value are read. Are compared (S1307). On the other hand, in the case of ON (in the route), the operation threshold value of the operation node is read (S1306), and the feature value and the action threshold value are compared (S1307). For example, if the feature amount is smaller than the threshold value, an operation node is identified from the smaller link destination (S1308). In this example, “less than” is a condition, but “below” may be a condition. On the other hand, if the feature amount is larger than the threshold value, the computing node is identified from the larger link destination (S1309). In this example, “more than” is a condition, but “greater than” may be a condition. If the identified next computation node is neither an OK termination nor an NG termination, the processing from S1302 is repeated. When the next specified operation node becomes an OK termination or an NG termination (S1310), a determination evaluation value is identified according to the type of termination (S1311). In the case of the OK termination, the judgment evaluation value is OK, and in the case of the NG termination, the judgment evaluation value is NG.

  FIG. 14 is a diagram illustrating an example (part 1) of the target block image path. A straight link line indicates the target block image path. In this example, the route is in the order of node number 2, node number 4, node number 5, and node number 7. The reference block path table in the state of FIG. 12 is assumed. Further, the middle block in the middle of the block images shown in FIG. 3 is the target block image.

  In node number 4, the median value of the pixel gradation is obtained as the feature amount. Since the reference block path flag of node number 4 is OFF (outside path), a standard threshold is used. The calculated median value “167” is compared with the standard threshold value “140”, and the node number 5 of the link destination on the larger side is reached.

  In node number 5, the mode value of the pixel gradation is obtained as the feature amount. Since the reference block path flag of node number 5 is OFF (outside path), a standard threshold is used. The calculated mode value “166” is compared with the standard threshold value “170”, and the node number 7 (NG termination) of the link destination on the smaller side is reached. Thereby, the judgment evaluation value becomes “NG”. The middle block in the middle of FIG. 3 has reached the determination result that there is a scratch while referring to the upper middle block.

  11, 12, and 14, the example in which the target block path does not overlap with the reference block path is shown. Next, in FIGS. 15, 16, and 17, the target block path is the reference block path. An overlapping example will be described.

  FIG. 15 is a diagram illustrating an example (part 2) of the reference block image path. FIG. 16 is a diagram illustrating an example (part 2) of the reference block path table. In this example, reference block image paths are formed in the order of node number 1, node number 3, node number 4, and node number 6. An example in which the upper right block in the block image shown in FIG. 3 is used as a reference block image is shown. Since the average value of the pixel gradation regarding the block on the upper right is “126”, it reaches the node number 4 which is the small link destination in the node number 3. In addition, the median pixel gradation for the upper right block is less than 140, and the processing ends at node number 6 (OK termination), which is the smaller link destination of node number 4.

  FIG. 17 is a diagram illustrating an example (part 2) of the target block image path. In this example, the route is in the order of node number 2, node number 4, node number 5, and node number 7. The reference block path table in the state of FIG. 16 is assumed. Further, the block on the right side of the middle stage among the block images shown in FIG. 3 is set as the target block image.

  In node number 4, the median value of the pixel gradation is obtained as the feature amount. At this time, since the reference block route flag of the node number 4 is ON (in the route), the action threshold is used. The calculated median value “135” is compared with the action threshold value “130”, and the node number 5 of the link destination on the larger side is reached.

  In node number 5, the mode value of the pixel gradation is obtained as the feature amount. Here, since the reference block path flag of node number 5 is OFF (outside path), a standard threshold is used. The calculated mode value “128” is compared with the standard threshold value “170”, and the node number 7 (NG termination) of the smaller link destination is reached. Thereby, the judgment evaluation value becomes “NG”. The middle right block of FIG. 3 has reached the determination result that there is a scratch while referring to the upper right block.

  Finally, the reproduction process (S204 in FIG. 2) will be described. FIG. 18 is a diagram showing a reproductive processing flow. First, an individual number determination unit (not shown) determines the number of individuals obtained by crossover processing, the number of individuals obtained by mutation processing, and the number of individuals obtained by life extension processing (S1801). The total number of these individuals is the total number of next-generation individuals. There are a method of using a predetermined constant and a method of determining each individual number based on a random number or the like.

  Crossover processing (S1802) for exchanging node information between two individuals by a crossover portion (not shown), and mutation processing (S1803) for changing part of node information by a mutation portion (not shown) Then, a life extension process (S1804) is performed by the life extension unit (not shown) to maintain the current individual structure. These processes are independent, and the order of the processes is not necessarily limited. When a next generation individual is obtained by these processes, an individual group update unit (not shown) performs an individual group update process (S1805). In this processing, the individual ID and the condition determination network individual structure relating to the current generation individual are deleted, and the condition determination network individual structure is stored in the individual group storage unit 102 in association with the individual ID for each next generation individual. That is, the current generation individual population is updated to the next generation individual population.

  The crossover process (S1802) will be described. FIG. 19 is a diagram showing a crossover process flow. The following processing is repeated until a predetermined number of individuals is obtained (S1901). Two individuals to be crossed are identified (S1902). Two individuals are selected from the selected individuals stored in the selected individual storage unit 108. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. Next, the range of nodes to be crossed is determined (S1903). There are a method of setting a single node as a crossing range, and a method of setting a plurality of nodes as a crossing range. There are a method for setting a continuous range such as node numbers 3 to 5 and a method for setting a non-continuous range such as node numbers 3, 5, and 7. Since the start and end are common to all individuals, they may be excluded from the crossover range. That is, only the intermediate node can be set as the crossover range. The crossover range is determined irregularly using random numbers or the like. There are two methods for making the crossover range common to the two individuals and for determining the crossover range independently. However, the number of nodes is matched even when determined independently. Node information is exchanged between individuals (S1904). That is, the data of the crossing range is copied to each other. In general, two crossed individuals are left in the next generation, but there is also a method of leaving only one of them. In this way, an individual obtained by crossover is specified, and the condition determination network individual structure is temporarily stored inside the reproductive part (S1905). These processes are repeated, and the process ends when a predetermined number of individuals is obtained (S1906).

  In this example, some information is exchanged between two individuals, but information may be exchanged between three or more individuals.

  The mutation process (S1803) will be described. FIG. 20 is a diagram showing a mutation process flow. The following process is repeated until a predetermined number of individuals is obtained (S2001). First, an individual to be mutated is identified (S2002). One individual is selected from the selected individuals stored in the selected individual storage unit 108. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. Data is changed with a predetermined mutation probability for all items of all nodes. Therefore, the following processing is repeated for each node information (S2003), and further, the following processing is repeated for each item of the node (S2004). When it corresponds with a predetermined probability (for example, 3%), it determines with changing (S2005). In the case of mutation, a mutation value within a predetermined numerical range for each item is determined (S2006). If it is a feature amount ID, it is randomly selected from predetermined feature amount IDs. If it is a link destination, for example, an intermediate node number larger than the node number is randomly selected. If it is a threshold value, it will select at random within the range of the assumed gradation value. Then, the item is rewritten to a mutation value (S2007). When all items of the record are repeated (S2008), and when all the record information is processed (S2009), an individual obtained by mutation is identified (S2010), and the condition judgment network is temporarily stored inside the reproductive part. Memorize the individual structure. These processes are repeated, and the process ends when a predetermined number of individuals is obtained (S2011).

  FIG. 21 is a diagram showing a life extension process flow. The following processing is repeated until a predetermined number of individuals is obtained (S2101). First, an individual whose life is to be extended is specified (S2102). One individual is selected from the individuals stored in the selected individual storage unit 107. Generally, the selection is made irregularly using random numbers or the like. There are methods that allow the same individual to be selected repeatedly, and methods that do not repeatedly select. As an individual whose life is to be extended, the condition determination network individual structure is temporarily stored inside the reproductive part. This process is repeated, and the process ends when a predetermined number of individuals is obtained (S2103).

Embodiment 2. FIG.
In the above-described example, the non-normal variable (tone value) indicating the characteristic is used as the feature amount and the threshold value as it is. Here, a description will be given of a mode in which a variable indicating characteristics is normalized dimensionlessly and a normal form feature amount and a normal form threshold value are used. In the following example, the gradation value is converted into a dimensionless amount in the range of 0 to 1 by normalization. As the conversion method, for example, proportional conversion or linear conversion can be considered.

  As shown in FIG. 22, the condition determination network individual structure uses a standard normal form threshold value and an action normal form threshold value as threshold values when receiving a standard threshold value and an action.

  In the initial individual input process (S201, FIG. 2), a condition determination network individual structure in which 0 to 1 values assumed to have normalized gradation values are set as the standard normal form threshold and the action normal form threshold is input.

  In the reference block image path determination process (S901, FIG. 10), after calculating the feature amount of the reference block image (S1004, FIG. 13), the calculated feature amount is normalized. Then, the feature amount normalized in S1005 is compared with the standard normal form threshold value. If the normalized feature value is on the small side, the process proceeds to S1006, and if the normalized feature value is on the large side, the process proceeds to S1007.

  Further, in the target block image path determination process (S902, FIG. 13), after calculating the feature amount of the target block image (S1303), the calculated feature amount is normalized. If it is determined in S1304 that the reference block path flag corresponding to the computation node is OFF (outside the path), the standard normal form threshold value of the computation node is read in S1305, and the feature amount calculated in S1307 is compared with the standard normal form threshold value. . On the other hand, if the reference block route flag corresponding to the computation node is ON (in the route) in the determination of S1304, the operation normal form threshold value of the operation node is read in S1306, and the feature amount and action normal form threshold value calculated in S1307 are obtained. Compare. If the normalized feature value is on the small side, the process proceeds to S1308. If the normalized feature value is on the large side, the process proceeds to S1309.

  In the mutation process (S1803, FIG. 20), when the item to be mutated is the standard normal form threshold value or the action normal form threshold value in S2006, the value of 0 to 1 that is assumed to have normalized the gradation value is changed. Value.

Embodiment 3 FIG.
An embodiment using an action width instead of an action threshold will be described. The action threshold value can be calculated from the standard threshold value and the action width. The action width corresponds to a difference obtained by subtracting the standard threshold value from the action threshold value. In the following example, −60 to +60 is set as the range of the action width.

  As shown in FIG. 23, the condition determination network individual structure has an action width instead of the action threshold.

  In the initial individual input process (S201, FIG. 2), a condition determination network individual structure in which a value of −60 to +60, which is assumed to be a difference between two gradation values, is set as an action width is input.

  In the target block image path determination process (S902, FIG. 13), when the reference block path flag corresponding to the calculation node is ON (within the path) in the determination of S1304, the standard threshold value and action width of the calculation node are read in S1306. The action threshold is obtained by adding the action width to the standard threshold. In step S1307, the feature amount is compared with the calculated action threshold value.

  In the mutation process (S1803, FIG. 20), when the item to be mutated is an action rate in S2006, a value of −60 to +60 that is assumed to be a difference between two gradation values is set as a mutation value.

Embodiment 4 FIG.
A mode in which a normal form action width is used instead of the action normal form threshold in the second embodiment will be described. The action normal form threshold value can be calculated from the standard normal form threshold value and the normal form action width. The normal form action width corresponds to a difference obtained by subtracting the standard normal form threshold value from the action normal form threshold value. In the following example, −0.3 to +0.3 is set as the range of the normal action width.

  As shown in FIG. 24, the condition determination network individual structure uses a standard normal form threshold value and a normal form action width.

  In the initial individual input process (S201, FIG. 2), a standard normal form threshold value of 0 to 1 that is assumed to have normalized the gradation value and a difference between the two normalized gradation values is assumed to be −0. A condition determination network individual structure in which a normal form action width of 3 to +0.3 is set is input.

  In the reference block image path determination process (S901, FIG. 10), after calculating the feature amount of the reference block image (S1004, FIG. 13), the calculated feature amount is normalized. Then, the feature amount normalized in S1005 is compared with the standard normal form threshold value. If the normalized feature value is on the small side, the process proceeds to S1006, and if the normalized feature value is on the large side, the process proceeds to S1007.

  Further, in the target block image path determination process (S902, FIG. 13), after calculating the feature amount of the target block image (S1303), the calculated feature amount is normalized. If it is determined in S1304 that the reference block path flag corresponding to the computation node is OFF (outside the path), the standard normal form threshold value of the computation node is read in S1305, and the feature amount calculated in S1307 is compared with the standard normal form threshold value. . On the other hand, if the reference block path flag corresponding to the computation node is ON (within the path) in the determination of S1304, the standard normal form threshold and normal form action width of the computation node are read in S1306, and the normal form is used as the standard normal form threshold. The action normal form threshold is obtained by adding the action width. Then, the feature amount normalized in S1307 is compared with the calculated action normal form threshold value. If the normalized feature value is on the small side, the process proceeds to S1308. If the normalized feature value is on the large side, the process proceeds to S1309.

  In the mutation process (S1803, FIG. 20), if the item to be mutated is the standard normal form threshold value in S2006, the value of 0 to 1 that is assumed to have normalized the gradation value is set as the mutation value. Further, when the item to be mutated is a normal form action width, a value of −0.3 to +0.3 assumed to be a difference between two normalized gradation values is set as a variation value.

Embodiment 5 FIG.
A mode of using the action rate instead of the action width will be described. The action threshold value can be calculated from the standard threshold value and the action rate. The action rate corresponds to a ratio obtained by dividing the action threshold value by the standard threshold value. In the following examples, 0.7 to 1.3 is set as the range of the operation rate.

  As shown in FIG. 25, the condition determination network individual structure uses an action rate instead of the aforementioned action width.

  In the initial individual input process (S201, FIG. 2), a condition determination network individual structure in which a ratio of two gradation values and a value of 0.7 to 1.3 assumed as an action rate is input.

  In the target block image path determination process (S902, FIG. 13), when the reference block path flag corresponding to the calculation node is ON (within the path) in the determination of S1304, the standard threshold value and the action rate of the calculation node are read in S1306. The action threshold is obtained by multiplying the standard threshold by the action rate. In step S1307, the feature amount is compared with the calculated action threshold value.

  In the mutation process (S1803, FIG. 20), when the item to be mutated is the action width in S2006, a value of 0.7 to 1.3 that is assumed to be a ratio of two gradation values is set as the mutation value. .

Embodiment 6 FIG.
A mode in which a normal-type action rate is used instead of the update-type action width in the fourth embodiment will be described. The action normal form threshold value can be calculated from the standard normal form threshold value and the normal form action rate. The normal form action rate corresponds to a ratio obtained by dividing the action normal form threshold value by the standard normal form threshold value. In the following example, 0.7 to 1.3 is set as the range of the normal form action rate.

  As shown in FIG. 26, the condition determination network individual structure uses a standard normal form threshold value and a normal form action rate.

  In the initial individual input process (S201, FIG. 2), a standard normal form threshold of 0 to 1 that is assumed to have normalized the gradation value and a ratio of two normalized gradation values that are assumed to be 0.7. A condition judgment network individual structure in which a normal form action rate of ˜1.3 is set is input.

  In the reference block image path determination process (S901, FIG. 10), after calculating the feature amount of the reference block image (S1004, FIG. 13), the calculated feature amount is normalized. Then, the feature amount normalized in S1005 is compared with the standard normal form threshold value. If the normalized feature value is on the small side, the process proceeds to S1006, and if the normalized feature value is on the large side, the process proceeds to S1007.

  Further, in the target block image path determination process (S902, FIG. 13), after calculating the feature amount of the target block image (S1303), the calculated feature amount is normalized. If it is determined in S1304 that the reference block path flag corresponding to the computation node is OFF (outside the path), the standard normal form threshold value of the computation node is read in S1305, and the feature amount calculated in S1307 is compared with the standard normal form threshold value. . On the other hand, if the reference block path flag corresponding to the computation node is ON (within the path) in the determination in S1304, the standard normal form threshold and normal form action rate of the computation node are read in S1306, and the normal form is used as the standard normal form threshold. The action normal form threshold value is obtained by multiplying the action rate. Then, the feature amount normalized in S1307 is compared with the calculated action normal form threshold value. If the normalized feature value is on the small side, the process proceeds to S1308. If the normalized feature value is on the large side, the process proceeds to S1309.

  In the mutation process (S1803, FIG. 20), if the item to be mutated is the standard normal form threshold value in S2006, the value of 0 to 1 that is assumed to have normalized the gradation value is set as the mutation value. Further, when the item to be mutated is a normal form action rate, a value of 0.7 to 1.3 which is assumed to be a ratio of two normalized gradation values is set as a mutation value.

Embodiment 7 FIG.
In the above-described example, an example in which the number of link destinations is two has been described, but three or more link destinations may be provided. In addition, the number of link destinations can be set for each node.

  FIG. 27 is a diagram illustrating a configuration of a condition determination network according to the seventh embodiment. Node number 13 uses two thresholds and sets two link destinations. For node number 14, four link destinations are set using three threshold values.

  FIG. 28 is a diagram illustrating an example of the condition determination network individual structure according to the seventh embodiment. The number of branches (same as the number of link destinations) is provided for each node. Also, the first to fourth link destinations are provided as the link destinations. The link destination corresponding to the number of branches is used. Ignore links that are larger than the number of branches. Moreover, the threshold of (number of branches-1) is used for both the standard threshold and the action threshold. Ignores thresholds greater than the number of branches. In this example, a normal form threshold value is shown, but the same applies to a non-normal form (variable as it is) threshold value.

  In the reference block image path determination process (S901, FIG. 10), subsequent to S1004, a node specifying process using a standard threshold is performed, and the process proceeds to S1008.

  FIG. 29 is a diagram showing a node specifying process flow based on the standard threshold. In this process, the link destination is determined based on the standard normal form threshold group according to the number of branches. For example, when the number of branches is 4, the first standard normal form threshold to the third standard normal form threshold are used, and if it is smaller than any threshold, the first link destination is specified, and other than the minimum threshold If it is smaller than the threshold value, the second link destination is specified, and if it is larger than the other threshold values except the maximum threshold value, the third link destination is specified. Specify the fourth link destination.

  Therefore, the following processing is repeated (number of branches−1) times (S2901). Then, the standard normal form threshold value is specified sequentially (S2902). That is, the first standard normal form threshold is used for the first time, the second standard normal form threshold is used for the second time, and the third standard normal form threshold is used for the third time. Then, the feature quantity is compared with the standard normal form threshold value (S2903), and when the feature quantity is on the large side, the large-side count is incremented (S2904). This is repeated (number of branches-1) times (S2905), and the number of large side is counted. Then, the (large-side count + 1) -th link destination is specified (S2906). That is, when the large number of times is 0, the first link destination is specified, when the large number of times is 1, the second link destination is specified, and when the large number of times is 2, the third link number is specified. Specify the link destination.

  Next, the target block image path determination process (S902, FIG. 13) will be described. FIG. 30 is a diagram showing a target block image path determination processing flow according to the seventh embodiment. Subsequent to S1303 in FIG. 13, the reference block path flag corresponding to the computation node is determined, and in the case of OFF (outside path), node identification processing based on the standard threshold value in FIG. 29 is performed (S3001). On the other hand, in the case of ON (within the route), node identification processing based on the action threshold shown in FIG.

  FIG. 31 is a diagram showing a node specifying process flow based on the action threshold. The link destination is determined based on the action normal form threshold group according to the number of branches. For example, when the number of branches is 4, the first action normal form threshold value to the third action normal form threshold value are used, and if it is smaller than any threshold value, the first link destination is specified, and the minimum threshold value is excluded. If it is smaller than the threshold value, the second link destination is specified, and if it is larger than the other threshold values except the maximum threshold value, the third link destination is specified. Specify the fourth link destination.

  Therefore, the following processing is repeated (number of branches-1) times (S3101). Then, the action normal form threshold value is identified sequentially (S3102). That is, the first action normal form threshold is used for the first time, the second action normal form threshold is used for the second time, and the third action normal form threshold is used for the third time. Then, the feature amount is compared with the action normal form threshold value (S3103), and when the feature amount is on the large side, the large number of times is incremented (S3104). This is repeated (number of branches-1) times (S3105), and the number of large side is counted. Then, the (large-side count + 1) -th link destination is specified (S3106). That is, when the large number of times is 0, the first link destination is specified, when the large number of times is 1, the second link destination is specified, and when the large number of times is 2, the third link number is specified. Specify the link destination.

Embodiment 8 FIG.
In the above-described example, the block image adjacent to the target block image is referred to, but a reference image other than the adjacent image may be used as sample data.

  Further, multidimensional data other than images (multidimensional vectors) may be used as determination target data. In that case, the same multidimensional data (multidimensional vector) is used as the reference data. The feature amount is calculated from the multidimensional data (multidimensional vector).

  Therefore, the target block image is an example of target multidimensional data (target multidimensional vector). The reference block image is an example of reference multidimensional data (reference multidimensional vector). The target block start edge is an example of the target multidimensional data start edge (target multidimensional vector start edge). The reference block start edge is an example of a reference multidimensional data start edge (reference multidimensional vector start edge). The reference block image path determination unit is an example of a reference multidimensional data path determination unit (reference multidimensional vector path determination unit). The target block image path determination unit is an example of a target multidimensional data path determination unit (target multidimensional vector path determination unit).

  The individual calculation unit 601 can be a condition determination network calculation device by itself. When it becomes a condition judgment network arithmetic unit, it has a condition judgment network individual structure storage unit for storing the condition judgment network individual structure, has a reference block image input unit for inputting a reference block image, and inputs a target block image A target block image input unit. Furthermore, a condition determination network individual structure input unit that inputs a condition determination network individual structure and stores it in the condition determination network individual structure storage unit may be provided. The reference block image input unit is an example of a reference multidimensional data input unit (reference multidimensional vector input unit). The target block image input unit is an example of a target multidimensional data input unit (target multidimensional vector input unit).

  The condition determination network generation system and the condition determination network calculation device are computers, and each element can execute processing by a program. Further, the program can be stored in a storage medium so that the computer can read the program from the storage medium.

101 initial individual input unit 102 individual population storage unit 103 fitness calculation unit 104 sample data set storage unit 105 target evaluation value storage unit 106 fitness storage unit 107 individual selection unit 108 selected individual storage unit 109 reproduction unit 110 learning individual output unit 601 Individual calculation unit 602 Correct answer determination unit 603 Correct answer counter 604 Correct answer rate calculation unit 801 Reference block image path determination unit 802 Reference block path table 803 Target block image path determination unit

Claims (8)

Target multidimensional data using a condition judgment network that links the start of target multidimensional data to be determined, the start of reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value A condition determination network operation device that determines an evaluation value for target multidimensional data based on data and reference multidimensional data, and has the following elements: (1) Each intermediate node A condition determination network individual structure storage unit for storing condition determination network individual structure information including node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values in the (2) reference multidimensional data from the start to the link The operation nodes are sequentially identified, and the feature amount for the reference multidimensional data is calculated for each operation node according to the feature amount ID. A reference multidimensional data path determination unit that generates a reference data path by comparing the threshold values and repeatedly selecting a link destination node based on the comparison result. (3) From the start of the target multidimensional data to the link The calculation nodes are sequentially identified along the line, the feature quantity for the target multidimensional data is calculated for each calculation node according to the feature quantity ID, and a threshold is selected based on whether the calculation node is a node on the reference data path. The target multidimensional data path determination that compares the feature value with the selected threshold and repeats the process of selecting the link destination node based on the comparison result, and specifies the evaluation value corresponding to the terminal when it reaches the terminal Department. The node information of the condition determination network individual structure information can specify a threshold value when receiving a standard threshold value and an action as a plurality of threshold values,
The target multidimensional data path determination unit selects a standard threshold value when the computation node is not a node on the reference data path, and sets a threshold value when the computation node is affected when the computation node is a node on the reference data path. 2. The condition judgment network computing device according to claim 1, wherein the condition judgment network computing device is selected.   3. The condition determination network computing device according to claim 1, wherein the target multidimensional data is image data to be determined, and the reference multidimensional data is image data to be referenced. A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation A determination network generation system, characterized by having the following elements: (1) Sample data set for storing sample data sets composed of target multidimensional data and reference multidimensional data for the number of samples Storage unit (2) Target evaluation value storage unit that stores target evaluation values corresponding to each sample data set (3) Node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values for each intermediate node Individual group storage unit for storing a group of condition determination network individual structure information (4) each condition determination network individual structure With respect to the information,
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
The fitness calculation unit for calculating the fitness of the condition determination network individual structure information based on the number of correct answers with respect to the number of samples (5) Based on the fitness, from the group of the condition determination network individual structure stored in the individual group storage unit Individual selection unit for selecting a part (6) Reproductive unit for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (7) Condition determination network individual structure based on fitness A learning individual output unit that selects and outputs. In order to define a condition judgment network that links the start of the target multidimensional data to be determined, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value Each node has a condition determination network individual structure storage unit for storing condition determination network individual structure information having node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values, and is referred to the target multidimensional data A condition determination network calculation method by a condition determination network calculation device for determining an evaluation value for target multidimensional data based on multidimensional data, the condition determination network calculation method having the following elements (1) The calculation nodes are sequentially identified along the link from the starting end of the reference multidimensional data, and the reference is made according to the feature amount ID for each calculation node. A reference multidimensional data path that calculates a feature quantity for multidimensional data, compares the feature quantity with one of the threshold values, and repeats the process of selecting a link destination node based on the comparison result to generate a reference data path Determination step (2) The calculation nodes are sequentially identified along the link from the starting end of the target multidimensional data, and the feature quantity for the target multidimensional data is calculated for each calculation node according to the feature quantity ID. Select a threshold based on whether or not it is a node, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and when it reaches the end, A target multidimensional data path determination step for identifying a corresponding evaluation value. In order to define a condition judgment network that links the start of the target multidimensional data to be determined, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value Each node has a condition determination network individual structure storage unit for storing condition determination network individual structure information having node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values, and is referred to the target multidimensional data A program for causing a computer that is a condition determination network computing device to determine an evaluation value for target multidimensional data based on multidimensional data to execute the following procedure: (1) Reference multidimensional data sequentially from the beginning to the end of the link A calculation node is specified, and a feature quantity for the reference multidimensional data is calculated according to the feature quantity ID for each calculation node. A reference multidimensional data path determination procedure for generating a reference data path by comparing the threshold values of the shifts and repeating the process of selecting a link destination node based on the comparison result. (2) From the start of the target multidimensional data to the link The calculation nodes are sequentially identified along the line, the feature quantity for the target multidimensional data is calculated for each calculation node according to the feature quantity ID, and a threshold is selected based on whether the calculation node is a node on the reference data path. The target multidimensional data path determination that compares the feature value with the selected threshold and repeats the process of selecting the link destination node based on the comparison result, and specifies the evaluation value corresponding to the terminal when it reaches the terminal procedure. A sample data set storage unit for storing sample data sets composed of target multidimensional data and reference multidimensional data for the number of samples;
A target evaluation value storage unit for storing a target evaluation value corresponding to each sample data set;
An individual group storage unit that stores a group of condition determination network individual structure information including node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values for each intermediate node;
A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation A condition determination network generation method by a determination network generation system, characterized by having the following elements: (1) For each condition determination network individual structure information,
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
A fitness calculation step for calculating the fitness of the condition determination network individual structure information based on the number of correct answers with respect to the number of samples (2) Based on the fitness, from the group of the condition determination network individual structure stored in the individual group storage unit Individual selection process for selecting a part (3) Reproductive process for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (4) Condition determination network individual structure based on fitness A learning individual output step of selecting and outputting. A sample data set storage unit for storing sample data sets composed of target multidimensional data and reference multidimensional data for the number of samples;
A target evaluation value storage unit for storing a target evaluation value corresponding to each sample data set;
An individual group storage unit that stores a group of condition determination network individual structure information including node information that can specify a feature amount ID, a plurality of link destinations, and a plurality of threshold values for each intermediate node;
A condition for generating a condition judgment network that links the start of the target multidimensional data to be judged, the start of the reference multidimensional data to be referenced, the intermediate node where the link destination branches, and the end corresponding to the evaluation value by evolutionary computation A program for causing a computer to be a determination network generation system to execute the following procedure (1) For each condition determination network individual structure information,
For each sample, obtain a sample data set and the corresponding target evaluation value,
The calculation nodes are sequentially identified from the starting end of the reference multidimensional data along the link, and for each calculation node, the feature amount for the reference multidimensional data of the sample is calculated according to the feature amount ID, and the feature amount is compared with one of the threshold values. , Repeat the process of selecting the link destination node based on the comparison result, generate the path of the reference data,
The calculation node is sequentially identified along the link from the start of the target multidimensional data, and the feature quantity for the target multidimensional data of the sample is calculated according to the feature quantity ID for each calculation node. The calculation node is a node on the reference data path. Select a threshold based on whether or not there is, compare the feature value with the selected threshold, repeat the process of selecting the link destination node based on the comparison result, and respond to the termination when it reaches the termination Specify the evaluation value as the evaluation value,
The correct answer is determined by matching the target evaluation value and the evaluation evaluation value,
Fitness calculation procedure for calculating fitness of condition determination network individual structure information based on the number of correct answers with respect to the number of samples (2) Based on fitness, from the group of condition determination network individual structures stored in the individual group storage unit Individual selection procedure for selecting a part (3) Reproductive procedure for generating a group of next-generation condition determination network individual structures based on the selected condition determination network individual structure (4) Condition determination network individual structure based on fitness Learning individual output procedure to select and output.