JP4660765B2 - Evolutionary image automatic classification apparatus, filter structure generation method, and program - Google Patents

Description

  The present invention relates to an image classification apparatus using a classifier (also referred to as a classifier) using a feature quantity group calculated from an image, and optimizes a filter string provided in a preceding stage of a feature quantity group calculation unit by evolutionary calculation. The present invention relates to a technique for improving the accuracy of image classification by obtaining converted image data suitable for classification by the classifier as an input image of a feature quantity group calculation unit.

  In recent years, automation by machines is rapidly progressing in industrial product production lines. In the inspection line, efforts have been made to automate the extraction of defective products by processing them with a computer using images. However, at present, a great deal of cost and time is required, such as a manual image processing by a person skilled in the image processing that sets conditions for recognizing wrinkles. Recently, learning methods that show high generalization ability have been devised in the field of pattern recognition, but even if they are applied, they still do not reach human recognition ability.

  Therefore, the purpose of the present invention is to obtain a discriminator capable of obtaining a high recognition result by combining the evolutionary image processing proposed by the laboratory to which the inventor belongs and these pattern recognition techniques. is there. Evolutionary image processing is the idea of formulating arbitrary given image processing as an optimization problem and solving the problem by an evolutionary calculation method.

  Pattern recognition is a process of matching an observed pattern with one of the concepts when the recognition target can be classified into several concepts. For example, the purpose of the recognition of numbers is to make the input pattern correspond to any of 10 numbers from 0 to 9.

In order to realize a machine that recognizes patterns, it is performed in two steps: feature extraction and learning / identification. In feature quantity extraction, some feature quantity is measured and extracted from the recognition target and digitized. Such feature quantities are collectively represented as a feature vector x = (x ₁ , x ₂ ,..., X _M ) ^T. If the selection of feature vectors is appropriate, feature vectors of the same class are similar to each other, and feature vectors of different classes are considered to be different from each other. From the feature vector obtained here, an identification method for determining which class the object belongs to is learned. For this purpose, it is necessary to first learn as a knowledge a stochastic correspondence between a feature vector and a class from a learning sample set with known class membership. Such learning is called supervised learning. Next, it is necessary to use a probabilistic knowledge about the correspondence between learned feature vectors and classes to determine which class the recognition target belongs to from the features of the given unknown recognition target. . At this time, it is desirable to minimize the probability of erroneous identification. This discrimination ability for unknown data is called generalization ability. A support vector machine (Support Vector Machine: SVM) was used as a pattern recognition technique.

  An image classification apparatus using SVM will be described. FIG. 25 is a diagram illustrating a configuration related to learning execution processing of a discriminator in the related art. L learning image classification data sets 2501 that are sets of learning image data 2502 and learning classification data 2502 are prepared, and a feature amount group calculation unit 2504 calculates a plurality of feature amounts from the learning image data 2502. Then, a process of inputting these feature quantity groups and learning classification data 2503 as a pair to the discriminator 2505 is performed for the L learning image classification data sets, and a learning execution process is performed.

  FIG. 26 is a diagram showing a configuration relating to classifier classification execution processing in the prior art. A feature amount group calculation unit 2504 calculates a feature amount group of the determination image data 2601, inputs the feature amount group to the discriminator 2505, and performs classification execution processing, thereby obtaining a classification result of the determination image data 2601.

  As described above, in image classification using SVM, there are two approaches for improving the correct answer rate of classification. One is to select a feature quantity suitable for the target classification as an input of the SVM, and the other is to extract the feature quantity more accurately from the image.

  Existing image analysis software can obtain more than 100 feature values from one image. Although all of these feature quantities can be input to the SVM for classification, in practice, selecting an appropriate input feature leads to an improvement in generalization ability. Conversely, the generalization ability does not necessarily improve as the feature quantity increases. Therefore, it is important to prepare feature quantities suitable for the target classification.

Also, when calculating feature values from an image, for example, if the target object cannot be accurately extracted due to noise in the image, the value of the feature value will not be accurate, and the classification will be adversely affected. Conceivable. In order to solve such a problem, it is conceivable to apply an image filter to the image in advance so that the feature quantity of the target object can be accurately extracted. However, the number of filters currently used in the field of image processing is enormous, and it is very difficult to manually determine which filter should be applied and in what order. The laboratory to which the inventor belongs has proposed evolutionary image processing that solves such problems using evolutionary computation methods such as GA. In evolutionary image processing, the structure of a filter is applied to GA or genetic programming (GP) to automatically construct target image processing.
JP 2001-134763 A JP 2005-309535 A

  In the present invention, with the goal of improving the correct answer rate in image classification, two methods are proposed with emphasis on extracting accurate feature values from images. In either case, an image filter is applied a plurality of times before an original image is input to a discriminator in SVM, and the structure is learned by GA.

The evolution type automatic image classification apparatus according to the present invention is:
Convert image data to be converted is generated by converting the image data to be converted, which is input to the feature amount group calculation unit that calculates the feature amount group to be input to the classifier by machine learning that determines the classification of the image from the image feature amount group Using the filter structure that defines the composition of the filter group to be used as an individual, select a part of the individual population according to the fitness from the individual population provided for each generation, and change the generation by reproductive processing of the individuals included in the selected individual group An evolutionary automatic image classification device that searches for a filter structure that is an individual that is an optimal solution or a practical solution by evolutionary calculation, and has the following elements.
(1) An individual group storage unit that stores an individual group for each generation with the filter structure as an individual. (2) The feature group calculation unit and the classifier are included, and each generation is stored in an individual group storage unit. For each filter structure that is included in an individual population, learning image data is generated for conversion according to the filter structure, and feature quantity groups are calculated from the generated converted image data for learning. Then, the classifier is trained to input the feature amount group of the converted image data for learning and output the classification data for learning which is the correct classification of the learning image data, and the evaluation image data as the conversion target A conversion image data for evaluation is generated according to the filter structure, a feature amount group is calculated from the generated conversion image data for evaluation, and a feature amount group of the conversion image data for evaluation is added to the learned discriminator. Input and output the classification result, calculate the correct answer rate where the classification result matches the classification data for evaluation which is the correct classification of the image data for evaluation, and use the calculated correct answer rate as the fitness value to determine the fitness of each individual Fitness calculation unit to be calculated (3) Individual selection unit for selecting a part of the individual population from the individual population according to the calculated fitness (4) Filter structure of individuals included in the selected individual group for each generation The reproductive part which produces | generates the filter structure used as a next-generation individual | organism | solid by reproductive processing, and memorize | stores the produced | generated filter structure group in an individual population memory | storage part as a next-generation individual population.

  The filter structure is a sequence of filter identification information indicating a multi-stage filter that sequentially inputs the output image data of the pre-stage filter and converts the image.

  The evolutionary image automatic classification apparatus further includes a learning individual output unit that outputs a filter structure of the fitness when the fitness that exceeds the fitness criterion that is a condition for a practical solution is exceeded. And

The filter structure generation method according to the present invention includes:
A classifier by machine learning that determines the classification of the image from the feature quantity group of the image, a feature quantity group calculation unit that calculates a feature quantity group to be input to the classifier, and converted image data to be input to the feature quantity group calculation unit A filter structure that defines the structure of a filter group that is generated by converting image data to be converted as an individual, and an individual group storage unit that stores an individual group for each generation, from an individual group provided for each generation An evolutionary model that searches for the filter structure, which is an individual solution that is an optimal solution or a practical solution, through evolutionary computation that selects a part of the population according to fitness and changes generations by reproductive processing of individuals included in the selected population. A filter structure generation method by an automatic image classification device, characterized by having the following elements.
(1) For each generation, for each filter structure that is an individual included in the individual group stored in the individual group storage unit, learning image data is generated as conversion target, and converted image data for learning is generated according to the filter structure. The feature amount group is calculated from the generated converted image data for learning, the feature amount group of the converted image data for learning is input, and the classification data for learning that is the correct classification of the learning image data is output. Learning the discriminator, generating conversion image data for evaluation according to the filter structure using the evaluation image data as a conversion target, calculating a feature amount group from the generated conversion image data for evaluation, Then, input the feature value group of the converted image data for evaluation, output the classification result, and calculate the correct answer rate that the classification result matches the classification data for evaluation which is the correct classification of the evaluation image data Then, the fitness calculation step of calculating the fitness of each individual using the calculated correct answer rate as the fitness (2) Individual selection step of selecting a part of the individual population from the individual population according to the calculated fitness for each generation ( 3) For each generation, a filter structure to be the next generation individual is generated from the filter structure of the individual included in the selected individual group by reproductive processing, and the generated filter structure group is stored in the individual population storage unit as the next generation individual population. Reproductive process to memorize.

  The filter structure generation method further includes a learning individual output step of outputting the filter structure of the fitness level when the calculated fitness level exceeds a fitness level that is a practical solution condition. .

The program according to the present invention is:
A classifier by machine learning that determines the classification of the image from the feature quantity group of the image, a feature quantity group calculation unit that calculates a feature quantity group to be input to the classifier, and converted image data to be input to the feature quantity group calculation unit A filter structure that defines the structure of a filter group that is generated by converting image data to be converted as an individual, and an individual group storage unit that stores an individual group for each generation, from an individual group provided for each generation An evolutionary model that searches for the filter structure, which is an individual solution that is an optimal solution or a practical solution, through evolutionary computation that selects a part of the population according to fitness and changes generations by reproductive processing of individuals included in the selected population. A computer that is an automatic image classification apparatus is caused to execute the following procedure.
(1) For each generation, for each filter structure that is an individual included in the individual group stored in the individual group storage unit, learning image data is generated as conversion target, and converted image data for learning is generated according to the filter structure. The feature amount group is calculated from the generated converted image data for learning, the feature amount group of the converted image data for learning is input, and the classification data for learning that is the correct classification of the learning image data is output. Learning the discriminator, generating conversion image data for evaluation according to the filter structure using the evaluation image data as a conversion target, calculating a feature amount group from the generated conversion image data for evaluation, Then, input the feature value group of the converted image data for evaluation, output the classification result, and calculate the correct answer rate that the classification result matches the classification data for evaluation which is the correct classification of the evaluation image data The fitness calculation procedure for calculating the fitness of each individual using the calculated correct answer rate as the fitness (2) Individual selection procedure for selecting a part of the individual population from the individual population according to the calculated fitness for each generation ( 3) For each generation, a filter structure to be the next generation individual is generated from the filter structure of the individual included in the selected individual group by reproductive processing, and the generated filter structure group is stored in the individual population storage unit as the next generation individual population. Reproductive procedure to remember.

  The program further causes the computer to execute a learning individual output procedure for outputting a filter structure of the fitness when the calculated fitness exceeds a fitness criterion that is a condition for a practical solution. .

  According to the present invention, in the image classification device using the classifier using the feature amount group calculated from the image, the feature amount group is optimized by performing evolutionary calculation on the filter row provided in the preceding stage of the feature amount group calculating unit. Converted image data suitable for classification by the classifier can be obtained as an input image of the calculation unit. Thereby, the accuracy of image classification can be improved.

Embodiment 1 FIG.
As a structure in which a filter array is applied before the classifier configuration in SVM, first, a structure as shown in FIGS. 1 and 2 was considered. That is, one specific filter row is made to correspond to all the feature quantities that are input to the SVM. (Hereinafter, this structure is expressed as a [1: all] structure.) Here, the filter train is a GA learning target.

  The discriminator is configured to determine the classification of the image from the feature amount group of the image by machine learning. And this apparatus calculates the fitness of a filter row | line | column through the process of two steps of the learning execution process and classification | category execution process by this discriminator. FIG. 1 is a diagram illustrating a configuration related to learning execution processing of a classifier. N learning image classification data sets 101, which are sets of learning image data 102 and learning classification data 103, are prepared. The learning image data 102 is image-converted by the filter array 104, and the feature amount group calculation unit 105 performs the image conversion. A plurality of different feature amounts are calculated from the converted image data for learning. Then, a process of inputting these feature quantity groups and the learning classification data 103 as a pair to the discriminator 2505 is performed on the N learning image classification data sets, and a learning execution process is performed. The learning classification data 103 indicates correct classification of the learning image data 102.

  Using the result learned by the discriminator 106 in this way, separately prepared image data for evaluation is classified, and the correct answer rate of the result is evaluated. FIG. 2 is a diagram illustrating a configuration related to classifier classification execution processing. M pieces of evaluation image classification data sets 201, which are sets of evaluation image data 202 and evaluation classification data 203, are prepared. The evaluation image data 202 is image-converted by the filter array 104, and the feature amount group calculation unit 105 converts the image data. A plurality of different feature amounts are calculated from the converted image data for evaluation. These feature amount groups are input to the discriminator 2505, and classification execution processing is performed. Then, it is determined whether or not the classification result matches the evaluation classification data 203. If they match, the answer is correct. If they do not match, the answer is incorrect. These processes are performed on the M image classification data sets for evaluation, and the correct answer rate is calculated. Then, this correct answer rate is treated as the GA adaptation rate, and an optimum or practical filter array is derived.

  The evaluation classification data 203 indicates the correct classification of the evaluation image data 102. The filter string is an example of a filter structure that defines the configuration of a filter group that is generated by converting image data to be converted.

  For example, this model uses a gamma correction filter to brighten an image that has become dark overall due to the environment in which the image was taken or the performance of the camera, or an image that has noise on it. It is expected that a structure for removing noise will be learned by a removal filter. If such a filter row is obtained, the object in the image becomes clearer, and it is considered that the classification by SVM is facilitated, leading to an improvement in the correct answer rate. Here, various structures of the filter row are conceivable. For example, a tree structure filter, a network filter, a hierarchical filter, and the like. These filter structures are highly expressive and can be expected to achieve high performance. Therefore, the filter row is not limited to one in series, and other filter structures may be used.

  For learning, a real value GA using a real value as it is instead of a bit string as a chromosome was used. A unique filter number was assigned to each of the prepared filter groups, and that number was used for the chromosome. That is, in this example, the filter sequence is a sequence of filter identification information indicating a multi-stage filter that sequentially inputs the output image data of the pre-filter and converts the image. The length of the filter is fixed, and a NOP filter (NonOperator) that does not modify the image is prepared so as to be pseudo variable. Further, as a filter to be used, a filter that abruptly converts an image is not appropriate because a feature amount of the image is likely to be lost. Therefore, a filter that does not change so rapidly was prepared.

An example of an image filter is shown.
Number Name: Contents F0 NOP: Do nothing (Non Operator)
F1 Inversion: Invert the gradation value F2 Maximum value: Replace with the maximum value of the gradation value near 8 of the target pixel F3 Minimum value: Replace with the minimum value of the gradation value near 8 of the target pixel F4 Median value: Attention F5 average value to be replaced with the median of the gradation values in the vicinity of the pixel 8: F6 range to be replaced with the average value of the gradation values in the vicinity of the pixel of interest F6 range: the maximum and minimum values of the gradation values in the vicinity of the pixel of interest 8 F7 variance replaced with difference: F8 replaced with variance in the vicinity of the pixel of interest F8 Prewitt: Previt filter F9 Sobel: Sobel filter F10 Vertical edge: F11 enhancing the vertical edge Horizontal edge: F12 enhancing the horizontal edge Laplacian: Laplacian filter F13 Sharpening: Sharpening by enhancing the edge F14 Original image difference: Difference in gradation value from the original image F15 LightPixel The dark pixels than binarization threshold by discriminant analysis 0
F16 DarkPixel: 255 pixels brighter than the binarization threshold by the discriminant analysis method
F17 Binary threshold distance: Replace with the difference between the gradation value of the target pixel and the binary threshold F18 Gamma correction (+): Gamma correction (γ = 2)
F19 Gamma correction (−): Gamma correction (γ = 0.5)
F20 Linear density conversion: Linear density conversion is performed between the minimum gradation value and the maximum gradation value. F21 Nonlinear density conversion: Non-linear density conversion is performed between the gradation values 0 and 255.
Number Name: Contents P0 Average gradation value: Average value of gradation values included in image P1 Standard deviation: Standard deviation of gradation values of all pixels P2 Maximum gradation value: Maximum gradation value included in image Value P3 Minimum gradation value: Minimum gradation value P4 included in image P4 Range: Maximum gradation value of all pixels minus minimum value P5 Probability within 1σ: Average gradation value ± standard deviation Proportion of number of pixels P6 3σ out-of-probability: average gradation value ± 3 × ratio of number of pixels not within standard deviation P7 First quartile: 1 / from the bottom when the gradation values are arranged in ascending order Gradation value P8 corresponding to position 4 Median: gradation value P9 corresponding to the central position when gradation values are arranged in ascending order Third quartile: 3/4 from the bottom when gradation values are arranged in ascending order Tone value P10 corresponding to the position Mode: Tone value P11 having the highest appearance frequency Distortion: Asymmetry of tone value distribution P12 Kurtosis: Degree of gradation value distribution P13 Discriminant analysis threshold value: Threshold value for binarization by discriminant analysis method P14 Low class variance: Tone value at binarization Dispersion P15 in the lower group Higher class dispersion: Dispersion P16 in the group with the higher gradation value in binarization Interclass dispersion: Dispersion P17 in the lower group in binarization Number of regions: Number of target areas after binarization P18 Average circumference: Average value of circumference of target area after binarization P19 Maximum circumference: Maximum value of circumference of target area after binarization P20 Average ferret diameter aspect ratio : Average value of aspect ratio of circumscribed rectangle of target area after binarization P21 Maximum ferret diameter aspect ratio: Maximum value of aspect ratio of circumscribed rectangle of target area after binarization P22 Total area area: After binarization Total value P23 of the area of the target area Average area area: of the target area after binarization Average area value P24 Maximum area area: Maximum area value of target area after binarization P25 Average circularity: Average degree of circularity (4πS / L2) of target area after binarization P26 Horizontal line difference ratio: 2 Ratio of the number of pixels with different gradation values in the upper and lower lines during horizontal line scan after binarization P27 Vertical line difference ratio: Ratio of the number of pixels with different gradation values in the left and right lines during vertical line scanning after binarization P28 Maximum angle : Maximum value P29 of the inclination angle of the target area after binarization Minimum angle: Minimum value P30 of the inclination angle of the target area after binarization Average angle: Average value of the inclination angle of the target area after binarization The evolution type automatic image classification apparatus according to the present invention will be described in detail. FIG. 3 is a diagram showing the configuration of the evolutionary image automatic classification apparatus. The evolution type automatic image classification apparatus includes a set data input unit 301, a learning image classification data set evaluation image classification data set storage unit 302, an initial individual input unit 303, an individual group storage unit 304, a fitness calculation unit 305, and a fitness. A storage unit 306, an individual selection unit 307, a selected individual storage unit 308, a reproduction unit 309, and a learning individual output unit 310 are provided.

  FIG. 4 is a diagram showing a flow of learning processing by the evolutionary calculation algorithm. First, learning image classification data set and evaluation image classification data set input processing (S401) is performed by the set data input unit 301, and the input learning image classification data set and evaluation image classification data set are used as learning image classification data. It is stored in the set evaluation image classification data set storage unit 302.

  Next, in the initial individual input process (S 402) by the initial individual input unit 303, a filter string to be an initial individual is input and stored in the individual group storage unit 304. The filter sequence may be randomly generated.

  In the generation change end determination process (S406), the generation change process is repeated until it is determined to end. Specifically, fitness calculation processing (S403) by the fitness calculation unit 305, individual selection processing (S404) by the individual selection unit 307, and reproduction processing (S405) by the reproduction unit 309 are repeated. The fitness calculation process (S403) for calculating the fitness of each individual (filter array) of the generation will be described later with reference to FIGS. In the individual selection processing (S404) by the individual selection unit 307, a part of the individual group is selected according to the fitness. For example, a predetermined number of individuals are selected in descending order of fitness. In the reproduction process (S405), a next generation individual (filter array) is generated based on the selected predetermined number of individuals. This process will be described later with reference to FIG.

  If it is determined in the generation change end determination process (S406) that the process ends, the fitness calculation unit 305 calculates the fitness for the final generation individual (filter string) (S407), and the learning individual output process (S408) by the learning individual output unit 310 ) To output the filter sequence and fitness. When the generation change end determination unit (not shown) determines that an optimal solution or a practical solution has been obtained, the determination is ended. For example, when the fitness level is constant even when the predetermined number of generations are changed, the optimal solution is determined and the process ends. Alternatively, a fitness standard that is a practical solution condition is set in advance, and the process ends when the calculated fitness exceeds the fitness standard.

  The fitness calculation processing (S403) by the fitness calculation unit 305 will be described in detail. FIG. 5 is a diagram illustrating a configuration of the fitness calculation unit. The fitness calculation unit 305 includes a classification learning unit 501, an evaluation classification unit 502, and a correct answer rate calculation unit 503.

  FIG. 6 is a diagram showing a fitness calculation processing flow. The following processing is repeated for each individual (filter row) (S601). First, classification learning processing (S602) by the classification learning unit 501 is performed. In this process, the filter train is provided in the preceding stage of the classifier to learn the classifier.

  FIG. 7 is a diagram illustrating a configuration of the classification learning unit. The classification learning unit 501 includes a filter array image conversion unit 701, a feature amount group calculation unit 702, a classifier learning execution unit 703, and a classifier learning data storage unit 704. The discriminator learning execution unit 703 and the discriminator learning data storage unit 704 constitute a part of the discriminator 106. That is, the classification learning unit 501 uses a configuration related to learning execution of the classifier 106.

  FIG. 8 is a diagram showing a classification learning process flow. The following processing is repeated for each learning image data (S801). First, filter row image conversion processing (S802) by the filter row image conversion unit 701 is performed.

  FIG. 9 is a diagram illustrating a configuration of the filter array image conversion unit. The filter row image conversion unit 701 includes a filter selection unit 901, filter execution units 902 to 904, and image data buffers 906 and 907.

  FIG. 10 is a diagram showing a filter row image conversion processing flow. First, the learning image data is written into the image data buffer A (S1001), the pointer of the image data buffer A is set as the input buffer pointer, and the pointer of the image data buffer B is set as the output buffer pointer (S1002). The input buffer pointer and the output buffer pointer are provided in the filter selection unit 901. Then, the following processing is sequentially repeated for each filter number in accordance with the filter column (S1003). The input buffer pointer and the output buffer pointer are set in the filter execution unit specified by the filter number, and the filter execution unit is activated (S1004). The filter execution unit reads the image data from the image data buffer specified by the input buffer pointer, converts the image, and writes the converted image into the image data buffer specified by the output buffer pointer. When the processing of the filter execution unit is completed, the input buffer pointer and the output buffer pointer are switched (S1005). When this process is performed for all the filter numbers (S1006), the converted image data is output from the image data buffer specified by the output buffer pointer (S1007).

  Next, a feature value group calculation process (S803 in FIG. 8) by the feature value group calculation unit 702 is performed. FIG. 11 is a diagram illustrating a configuration of the feature amount group calculation unit. A plurality of feature amount calculation units 1101 to 1103 are provided. These feature amount calculation units correspond to predetermined feature amounts determined as the specifications of the input of the classifier.

  FIG. 12 is a diagram showing a feature quantity group calculation processing flow. The converted image data is input to each feature amount calculation unit to calculate the feature amount. In this example, a first feature value calculation process (S1201) by the first feature value calculation unit 1101, a second feature value calculation process (S1202) by the second feature value calculation unit 1102, and a third feature value calculation unit 1103 by the third feature value calculation unit 1103. A feature amount calculation process (S1203) is performed.

  Then, at the time when the feature amount group is calculated for all the learning image data (S804 in FIG. 8), the classifier learning execution process (S805) by the classifier learning execution unit 703 is performed. N pairs of feature quantity groups and classification data for learning are input, and learning as a classifier is performed. The learning result is stored in the discriminator learning data storage unit 704 as discriminator learning data (sometimes referred to as a classification value in SVM). This data is used in the classifier classification execution process described later.

  Next, evaluation classification processing (S603 in FIG. 6) by the evaluation classification unit 502 is performed. In this process, the filter row is provided before the learned discriminator, and the evaluation image is classified by the discriminator.

  FIG. 13 is a diagram illustrating a configuration of the evaluation classification unit. The evaluation classification unit 502 includes a filter array image conversion unit 701, a feature amount group calculation unit 702, a classifier classification execution unit 1301, and a classifier learning data storage unit 704. The classifier classification execution unit 1301 and the classifier learning data storage unit 704 constitute part of the classifier 106. That is, the evaluation classification unit 502 uses the configuration related to the classification execution of the classifier 106.

  FIG. 14 is a diagram showing an evaluation classification process flow. The following processing is repeated for each evaluation image data (S1401 to S1404), and a classification result is obtained for each. Similarly to the classification learning process, the filter column image conversion unit 701 performs filter column image conversion processing (S1402), and then the feature amount group calculation unit 702 performs feature amount group calculation processing (S1403). Then, a classifier classification execution process (S1404) by the classifier classification execution unit 1301 is performed. In this processing, a feature amount group of converted image data for evaluation is input, and a classification result is calculated using classifier learning data. Then, the process ends when all the evaluation image data has been processed (S1405).

  Next, a correct answer rate calculation process (S604 in FIG. 6) is performed by the correct answer rate calculator 503. In this process, the correct answer rate of the classification result for the evaluation image is calculated.

  FIG. 15 is a diagram showing a correct answer rate calculation processing flow. The following processing is repeated for each classification result (S1501). It is determined whether the classification result matches the evaluation classification data corresponding to the original evaluation image data (S1502). If they match, the number of correct answers (initial value = 0) is incremented and counted up. (S1503). When all the classification results are processed (S1504), the number of correct answers is divided by the number of classification results (M) to obtain the quotient as the correct answer rate (S1505).

The fitness calculation processing in FIG. 6 ends when the above processing is performed for all individuals. (S605 in FIG. 6)
The reproduction process (S405 in FIG. 4) by the above-described reproduction unit 309 will be described. FIG. 16 is a diagram showing a reproductive processing flow. The number of individuals obtained by the crossover process, the number of individuals obtained by the mutation process, and the number of individuals obtained by the life extension process are determined (S1601). In the crossover process (S1602), crossover is performed to obtain new individuals for the number of individuals. Crossover is performed by selecting a plurality of individuals from the selected individual storage unit 308, setting a crossover point of the filter row that is the individual, and a row of filter numbers between the crossover points and a filter number row (chromosome between the crossover points). Column) are exchanged with each other to generate a new filter column. In the mutation process (S1603), mutation is performed to obtain new individuals for the number of individuals. Mutation is performed by selecting an individual from the selected individual storage unit 308, selecting a filter number (chromosome) of a single or substring to be mutated, and replacing it with another filter. Other filters may be determined randomly. In the life extension process (S1604), individuals corresponding to the number of individuals are selected from the selected individual storage unit 308 and used as the next generation individuals as they are. The new individual group obtained in this way is written into the individual group storage unit 304 by the individual group update process (S1605). That is, the current generation population is updated to the next generation population.

Embodiment 2. FIG.
In the above-mentioned [1: all] structure, the structure of learning one filter row is ignored while ignoring the meaning of the feature quantity corresponding to the input of the SVM. On the other hand, the structures as shown in FIGS. 17 to 19 are considered so that values that can be easily classified can be input to the respective feature amounts. That is, one filter row is associated with each feature quantity that is an input of the SVM. (Hereinafter, this structure is expressed as a [1: 1] structure.) Similar to the [1: all] structure, the filter train is a GA learning target.

  In addition to the above-described improvements, this structure is expected to provide a filter row in which a certain feature value input to the SVM is clarified for each type of image. For example, when the desired image classification is an image with many edges and an image with no edges, the classification can be facilitated by associating the edge detection filter with the feature amount of the area. Here, it is unlikely that applying an edge detection filter to all feature quantities will yield good results, and it was considered useful to prepare a filter for each feature quantity. On the other hand, since a filter row is prepared for each feature quantity, it is conceivable that the search space becomes enormous in optimization by GA, and it is difficult to obtain an optimal solution or it takes time.

  The learning of the filter row was performed by connecting the respective filter row portions to one and considering it as one filter row, and learning was performed in the same manner as in the [1: all] structure.

  FIG. 17 is a diagram illustrating a configuration of a filter array according to the second embodiment. As shown in the figure, one filter row is treated as a series of a plurality of filter row portions 1701 to 1703.

  FIG. 18 is a diagram illustrating a configuration relating to learning execution processing of the classifier according to the second embodiment. The first feature amount converted image data converted by the first feature amount filter row portion 1701 is used for the first feature amount calculation by the feature amount group calculation unit 105, and similarly converted by the second feature amount filter row portion 1702. The second feature quantity converted image data is used for the second feature quantity calculation by the feature quantity group calculation unit 105, and the third feature quantity converted image data converted by the third feature quantity filter row portion 1703 is used as the feature quantity. It is used for the third feature amount calculation by the group calculation unit 105.

  FIG. 19 is a diagram illustrating a configuration related to classification execution processing of the classifier according to the second embodiment. Also in the classification execution process, the converted image data for each feature amount (the first feature amount conversion image data, the second feature amount conversion image data, and the third feature amount conversion image data) as in the learning execution processing. Used to calculate each feature amount.

  FIG. 20 is a diagram illustrating a configuration of the classification learning unit according to the second embodiment. The input to the filter row image conversion unit 701 is a plurality of feature amount-specific filter row portions (first feature amount filter row portion, second feature amount filter row portion, third feature amount filter row portion) and Become.

  FIG. 21 is a diagram illustrating a configuration of a filter array image conversion unit according to the second embodiment. FIG. 22 is a diagram showing a filter array image conversion processing flow according to the second embodiment. The following processing is repeated for each feature quantity filter row portion (S2201). In step S2204, the following processing is repeated for each filter number in order from the feature amount filter array portion. Every time the loop processing corresponding to the filter column portion classified by feature amount is completed (S2201 to S2207), the converted image data classified by feature amount is output from the image data buffer specified by the output buffer pointer (S2208). Then, the process is terminated when all the filter columns for each feature amount are processed (S2209).

FIG. 23 is a diagram illustrating a configuration of a feature amount group calculation unit according to the second embodiment. Each feature amount calculation unit calculates each feature amount using the converted image data classified by feature amount converted exclusively.
FIG. 24 is a diagram illustrating a configuration of the evaluation classification unit according to the second embodiment. Similar to the configuration of the classification learning unit, the input to the filter array image conversion unit 701 is a plurality of feature amount-specific filter arrays.

  In the reproductive process, the crossover process, the mutation process, and the life extension process are performed for the entire filter row in the same manner as described above, and the crossover process, the mutation process, and the life extension process are performed for each filter row portion by feature amount. The method is also effective.

  In the above example, the classifier is not limited to SVM, and other clustering may be used. For example, another machine learning classifier such as a neural network may be used. In that case, the learning neural network is trained by the error back-propagation method so as to input the feature amount group and output the classification data, and the resultant connection weight becomes the discriminator learning data. Then, a feature group is input to the classification neural network having the same configuration as the learning neural network, and the neural network calculation processing is performed according to the coupling weight of the discriminator learning data storage unit, and the classification result is obtained as a processing result. Output.

  The evolutionary automatic image classification apparatus is a computer, 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. In the computer, an arithmetic device, a data storage device, and a memory are connected to a bus. The data storage device is, for example, a ROM (Read Only Memory) or a hard disk. The memory is usually a RAM (Random Access Memory). The program is normally stored in the data storage device, and is loaded into the memory and sequentially read into the arithmetic device for processing.

It is a figure which shows the structure which concerns on the learning execution process of a discriminator. It is a figure which shows the structure which concerns on the classification | category execution process of a discriminator. It is a figure which shows the structure of an evolution type automatic image classification device. It is a figure which shows the flow of the learning process by an evolution calculation algorithm. It is a figure which shows the structure of a fitness calculation part. It is a figure which shows a fitness calculation processing flow. It is a figure which shows the structure of a classification | category learning part. It is a figure which shows a classification | category learning process flow. It is a figure which shows the structure of a filter row | line | column image conversion part. It is a figure which shows a filter sequence image conversion processing flow. It is a figure which shows the structure of a feature-value group calculation part. It is a figure which shows the feature-value group calculation processing flow. It is a figure which shows the structure of an evaluation classification | category part. It is a figure which shows an evaluation classification process flow. It is a figure which shows a correct rate calculation process flow. It is a figure which shows a reproductive process flow. 6 is a diagram illustrating a configuration of a filter array according to Embodiment 2. FIG. It is a figure which shows the structure which concerns on the learning execution process of the discriminator which concerns on Embodiment 2. FIG. It is a figure which shows the structure which concerns on the classification | category execution process of the discriminator which concerns on Embodiment 2. FIG. 6 is a diagram illustrating a configuration of a classification learning unit according to Embodiment 2. FIG. 6 is a diagram illustrating a configuration of a filter array image conversion unit according to Embodiment 2. FIG. FIG. 10 is a diagram showing a filter array image conversion processing flow according to the second embodiment. FIG. 10 is a diagram illustrating a configuration of a feature amount group calculation unit according to Embodiment 2. It is a figure which shows the structure of the evaluation classification | category part which concerns on Embodiment 2. FIG. It is a figure which shows the structure which concerns on the learning execution process of the discriminator in a prior art. It is a figure which shows the structure which concerns on the classification | category classification | category execution process in a prior art.

Explanation of symbols

101 learning image classification data set, 102 learning image data, 103 learning classification data, 104 filter train, 105 feature quantity group calculation unit, 106 classifier 201, evaluation image classification data set, 202 evaluation image data, 203 Classification data for evaluation, 204 correct answer rate calculation unit, 301 set data input unit 302, learning image classification data set evaluation image classification data set storage unit, 303 initial individual input unit, 304 individual group storage unit, 305 fitness calculation unit , 306 fitness storage unit, 307 individual selection unit, 308 selected individual storage unit, 309 reproduction unit, 310 learning individual output unit, 501 classification learning unit, 502 evaluation classification unit, 503 correct answer rate calculation unit, 701 filter sequence image conversion unit 702, feature quantity group calculation unit, 703 classifier learning execution unit, 704 classifier learning data storage unit 901 Filter selection unit, 902 to 904 Filter execution unit, 905 Image data buffer A, 906 Image data buffer B, 1101 First feature value calculation unit, 1102 Second feature value calculation unit, 1103 Third feature value calculation unit, 1301 Identification Device classification execution unit, 1701 first feature amount filter row portion, 1702 second feature amount filter row portion, 1703 third feature amount filter row portion, 2501 learning image classification data set, 2502 learning image data, 2503 Classification data for learning, 2504 feature quantity group calculation unit, 2505 classifier, 2601 determination image data.

Claims (9)

Convert image data to be converted is generated by converting the image data to be converted, which is input to the feature amount group calculation unit that calculates the feature amount group to be input to the classifier by machine learning that determines the classification of the image from the image feature amount group Using the filter structure that defines the composition of the filter group to be used as an individual, select a part of the individual population according to the fitness from the individual population provided for each generation, and change the generation by reproductive processing of the individuals included in the selected individual group An evolutionary automatic image classification device that searches for a filter structure that is an individual that is an optimal solution or a practical solution by evolutionary computation, and has the following elements: (1) Filter An individual group storage unit that stores an individual group for each generation with the structure as an individual (2) has the feature value group calculation unit and the discriminator, and stores them in the individual group storage unit for each generation For the filter structure that is each individual included in the body group, learning image data is generated according to the filter structure with the learning image data as a conversion target, and a feature amount group is calculated from the generated converted image data for learning. The classifier is trained to input a feature amount group of converted image data for learning and output learning classification data that is a correct classification of the image data for learning, and the filter using the evaluation image data as a conversion target Generate converted image data for evaluation according to the structure, calculate a feature amount group from the generated converted image data for evaluation, and input the feature amount group of the converted image data for evaluation into the learned discriminator to classify The result is output, and the correct answer rate that the classification result matches with the classification data for evaluation that is the correct classification of the evaluation image data is calculated. Fitness calculation unit for calculating the fitness (3) Individual selection unit for selecting a part of the individual population from the individual population according to the calculated fitness (4) Individuals included in the selected individual group for each generation The reproductive part which produces | generates the filter structure used as a next generation individual from the filter structure of this by reproductive processing, and memorize | stores the produced | generated filter structure group as a next generation individual population in an individual population memory | storage part.   2. The evolution type automatic image classification apparatus according to claim 1, wherein the filter structure is a sequence of filter identification information indicating a multi-stage filter for sequentially inputting the output image data of the preceding filter and converting the image.   The evolutionary image automatic classification apparatus further includes a learning individual output unit that outputs a filter structure of the fitness when the fitness that exceeds the fitness criterion that is a condition for a practical solution is exceeded. The evolution type automatic image classification apparatus according to claim 1. A classifier by machine learning that determines the classification of the image from the feature quantity group of the image, a feature quantity group calculation unit that calculates a feature quantity group to be input to the classifier, and converted image data to be input to the feature quantity group calculation unit A filter structure that defines the structure of a filter group that is generated by converting image data to be converted as an individual, and an individual group storage unit that stores an individual group for each generation, from an individual group provided for each generation An evolutionary model that searches for the filter structure, which is an individual that is the optimal solution or practical solution, through evolutionary computation that selects a part of the population according to fitness and changes generations by reproductive processing of individuals included in the selected population A filter structure generation method using an automatic image classification apparatus, characterized by having the following elements: (1) Individual population stored in an individual population storage unit for each generation For each filter structure that is included, learning image data is converted into learning object, conversion image data for learning is generated according to the filter structure, a feature amount group is calculated from the generated conversion image data for learning, and learning The classifier is trained to input a feature amount group of converted image data and output learning classification data that is a correct classification of the learning image data, and evaluation image data is subjected to evaluation according to the filter structure. Conversion image data is generated, a feature amount group is calculated from the generated conversion image data for evaluation, and the classification amount is output by inputting the feature amount group of the conversion image data for evaluation into the learned classifier And calculate the correct answer rate that matches the classification data for evaluation, which is the correct classification of the image data for evaluation, and use the calculated correct answer rate as the fitness value to determine the fitness of each individual. Fitness calculation step to be performed (2) individual selection step of selecting a part of the individual population from the individual population according to the calculated fitness (3) filter structure of individuals included in the selected individual population for each generation The reproductive process which produces | generates the filter structure used as the next generation individual | organism | solid by reproductive processing, and memorize | stores the produced | generated filter structure group in the individual population memory | storage part as a next generation individual population.   5. The filter structure generation method according to claim 4, wherein the filter structure is a sequence of filter identification information indicating a multi-stage filter for sequentially inputting the output image data of the preceding filter and converting the image.   The filter structure generation method further includes a learning individual output step of outputting the filter structure of the fitness level when the calculated fitness level exceeds a fitness level that is a practical solution condition. The filter structure generation method according to claim 4. A classifier by machine learning that determines the classification of the image from the feature quantity group of the image, a feature quantity group calculation unit that calculates a feature quantity group to be input to the classifier, and converted image data to be input to the feature quantity group calculation unit A filter structure that defines the structure of a filter group that is generated by converting image data to be converted as an individual, and an individual group storage unit that stores an individual group for each generation, from an individual group provided for each generation An evolutionary model that searches for the filter structure, which is an individual that is the optimal solution or practical solution, through evolutionary computation that selects a part of the population according to fitness and changes generations by reproductive processing of individuals included in the selected population A program for causing a computer to be an automatic image classification apparatus to execute the following procedure (1) A filter that is each individual included in the individual group stored in the individual group storage unit for each generation The learning image data is converted into the conversion target data according to the filter structure, the feature amount group is calculated from the generated conversion image data for learning, and the feature amount of the converted conversion image data for learning The classifier is trained to input a group and output learning classification data that is the correct classification of the learning image data, and generates evaluation converted image data according to the filter structure with the evaluation image data as a conversion target. Then, a feature value group is calculated from the generated converted image data for evaluation, and the learned discriminator is input with the feature value group of the converted image data for evaluation and outputs a classification result. The fitness calculation procedure for calculating the fitness of each individual using the calculated correct response rate as the fitness, and calculating the fitness rate corresponding to the classification data for evaluation, which is the correct answer classification of the image data (2) Individual selection procedure for selecting a part of the individual population from the individual population according to the calculated fitness (3) For each generation, the filter structure to be the next generation individual from the filter structure of the individual included in the selected individual group A reproductive procedure that is generated by reproductive processing and stores the generated filter structure group in the individual population storage unit as the next generation individual population.   8. The program according to claim 7, wherein the filter structure is a sequence of filter identification information indicating a multi-stage filter for sequentially inputting the output image data of the preceding-stage filter and converting the image. The program further causes the computer to execute a learning individual output procedure for outputting a filter structure of the fitness when the calculated fitness exceeds a fitness criterion that is a condition for a practical solution. The program according to claim 7.

Images