JP4852086B2 - Pattern recognition device - Google Patents

Description

  The present invention relates to a pattern recognition apparatus that performs pattern recognition of characters and the like.

  The parameters of the feature extraction unit in a pattern recognition apparatus that recognizes patterns such as conventional characters are designed manually. However, for example, when it is necessary to change the feature extraction method and parameters for each group of similar characters in character recognition, the number of methods is enormous, so it is possible to actually create the parameters of the feature extraction unit manually. Have difficulty. For this reason, techniques for automatically determining feature extraction parameters using learning are mainly known in the field of speech recognition. For example, there are the following open technologies (Non-Patent Documents 1 to 4).

  The technologies disclosed in Non-Patent Documents 1 and 2 use a speech signal as an input, and obtain parameters of a filter bank and a lifter when extracting features by a probability drop method. However, these techniques cannot be used for feature extraction for character recognition.

  Non-Patent Documents 3 and 4 obtain a feature vector by performing linear transformation on an input vector. The calculation formula representing the update rule of the feature transformation matrix is different from the calculation formula for filter learning in the present invention. However, the expression itself has the same content. However, since the parts to which the learning rule is applied are different between the conventional technique and the present invention, and the purpose of learning is completely different, the present invention and the feature transformation matrix learning are essentially different systems. What the present invention is learning is a parameter of a filter when a feature vector is created from an input image, whereas feature transformation matrix learning of the conventional method further linearizes a feature vector obtained from an input image. It is learning with respect to a linear transformation matrix at the time of transformation. The feature transformation matrix learning does not have the concept of partial regions introduced in the present invention, and merely provides general learning rules for general vector-to-vector linear transformation. Further, in the present invention, a nonlinear suppress function is introduced into the process of creating the feature vector, and the feature extraction does not have to be a linear transformation. In short, the conventional method cannot learn feature extraction itself.

An object of the present invention is to recognize a similar pattern with high accuracy. The following methods have been proposed as similar character recognition methods (see Patent Documents 1 to 3). In these documents, a method for recognizing or checking a different part of similar characters is disclosed.
Feature Extraction Based on Minimum Classification Error / Generalized Probabilistic Descent Method ”, A. Biem et.al., Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, Vol.2, pp275-278, (1993) “Filter Bank Design Based on Discriminative Feature Extraction”, A. Biem et.al., Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, Vol.1, pp485-488, (1994) “Minimum Classification Error Training Algorithm for Feature Extractor and Pattern Classifier in Speech Recognition”, KK paliwal et.al, EUROSPEECH'95, vol.1, pp541-544, (1995) "Character selection type character recognition by minimum classification error learning" Kawamura et al., IEICE Transactions D-II, Vol. J81-D-II, No. 12, pp. 2749-2756, (1998) JP 2004-185264 A JP 2003-162688 A JP-A-11-259599

  In the conventional known method, since recognition is performed by paying attention to only a part of the input image, there is a problem that the recognition cannot be performed accurately if the position to be focused is inappropriate. In addition, the feature extraction parameters cannot be automatically determined by learning.

  The present invention is to solve such problems, and a pattern capable of improving the recognition accuracy of similar characters by performing feature extraction with an increased focus on the parts important for the recognition of the input image. The object is to provide a recognition device.

  In order to solve the above-described problems, the pattern recognition apparatus of the present invention calculates the feature amount of each of a plurality of partial areas that are numerically nonuniformly set in the image area with respect to the input pattern image. Output from the feature vector generation unit, a feature vector generation unit that generates a feature vector of a pattern input from the feature amounts of each of the plurality of partial regions calculated by the partial region feature amount calculation unit, and a feature vector generation unit And a pattern identification unit for identifying an input pattern using the feature vector. In the present invention, the plurality of partial areas are set so as to be arranged numerically densely in a part of the image area as compared with other parts. By adopting the configuration in this way, it is possible to perform feature extraction with an increased focus on the parts important for the recognition of the input image, and the recognition accuracy of similar characters is improved.

  In the present invention, the partial region feature amount calculation unit is provided for each similar recognition target group. Also, by providing a plurality of partial area feature quantity calculation units for each similar recognition target group, feature extraction parameters corresponding to features that reflect personal habits appearing in handwritten characters are set in the partial area feature quantity calculation units, respectively. Recognition accuracy is improved.

  Further, the plurality of partial regions may be formed by dividing the entire region by a plurality of lines in the biaxial direction. According to this configuration, the feature extraction parameter can be easily changed by changing the positions of the plurality of lines in the biaxial direction, and the process for generating the feature extraction parameter by learning is simplified.

  Furthermore, the pattern recognition apparatus of the present invention evaluates a result obtained by recognizing a learning pattern each time changing one or a plurality of partial areas, thereby setting a plurality of patterns set in the partial area feature amount calculation unit. A learning unit that selects the partial area may be further included. Thereby, a pattern recognition apparatus capable of automatically generating feature extraction parameters by learning can be provided.

  Further, the feature vector generation unit is inputted from the feature amount obtained by multiplying the feature amount of each of the plurality of partial regions calculated by the partial region feature amount calculation unit by the weight value determined for each partial region. It is assumed that a feature vector of a pattern is generated, and the weight value for each partial region is a loss function for obtaining a loss value of a difference or similarity with a standard pattern obtained by recognizing a learning pattern. May be changed by evaluating the result obtained by recognizing the differentiated value and the learning pattern. Thereby, the feature extraction parameters can be automatically generated by learning.

  Further, the following means can be used as a modification of the above means. The feature vector generation unit is input from the feature value corresponding to the partial region calculated by the product value of the pixel value in the partial region or the feature value corresponding to the pixel and the weight corresponding to the pixel in the partial region feature amount calculation unit. The feature vector of the generated pattern is generated, and the weight value corresponding to the pixel for each partial region is used to determine the difference value or the similarity loss value from the standard pattern obtained by recognizing the learning pattern. The loss function may be changed by differentiating with a variable representing the weight value and evaluating the result obtained by recognizing the differentiated value and the learning pattern.

  According to the present invention, it is possible to perform feature extraction with an increased focus on a portion that is important for recognition of an input image, and improve recognition accuracy of similar characters. Also, feature extraction parameters can be automatically generated by learning.

  Embodiments of the present invention will be described below.

(First embodiment)
FIG. 1 is a block diagram showing the configuration of a pattern recognition apparatus according to the first embodiment of the present invention. As shown in the figure, the pattern recognition apparatus 100 of this embodiment includes a primary recognition unit 110 and a secondary recognition unit 120.

The primary recognition unit 110 includes a feature extraction unit 111, a pattern identification unit 112, and a dictionary 113. The feature extraction unit 111 extracts a feature of the input pattern by a known method such as a density pattern method or a weighted direction histogram method to generate a feature vector. The pattern identification unit 112 collates the feature vector input from the feature extraction unit 111 with the feature vector of the standard pattern of the detection target character type stored in the dictionary 113 to determine one or more character candidates having high similarity. Then, the determination result is given to the secondary recognition unit 120 as the output of the primary recognition unit 110.

  The secondary recognition unit 120 includes a feature extraction unit 121, a pattern identification unit 122, and a dictionary 123. The feature extraction unit 121 further includes a plurality of partial region feature amount calculation units 124-x (124-0, 124-1,...), A switching unit 125, and a feature vector generation unit 126.

  The plurality of partial region feature amount calculation units 124-x are provided for each group of similar characters. One of the plurality of partial region feature quantity calculation units 124-x is selected by the switching unit 125, and the selected partial region feature quantity calculation unit 124-x is set in advance for the input pattern input to the primary recognition unit 110. The feature amount for each of the plurality of partial areas is calculated.

  The calculation of the feature amount in one partial region feature amount calculation unit 124-x is performed as follows. FIG. 2 is a diagram illustrating an example of an input pattern image, and FIG. 3 is a diagram illustrating an example of a plurality of partial regions that are set numerically non-uniformly in the entire region of the pattern image. Here, it is assumed that K partial areas are Ai (i = 0,..., K−1). K is an arbitrarily determined value. Each of these partial areas Ai is an area that occupies a part of the entire pattern image, and may be partly separated from each other, or may be arranged so as to partially overlap each other.

  FIG. 3 shows the arrangement of the partial areas Ai set in the partial area feature amount calculation unit 124-x corresponding to the group of similar characters including the character “B”. In addition, “ho”, “po”, “ha”, “ba”, “pa”, and the like belong to the group of similar characters including the characters “bo”. In these similar characters, the upper right part is an important part for recognition. Therefore, the partial areas Ai are arranged relatively densely in the upper right part, and are arranged relatively sparsely in the other parts. Here, densely arranging the partial areas Ai means arranging a small number of small partial areas, and sparsely arranging means arranging a small number of large partial areas. . Such partial region setting contents are set in the partial region feature amount calculation unit 124-x as a feature extraction parameter.

  Further, weight values Wij (j = 0,..., M−1) are respectively determined for the pixels Pij (j = 0,..., M−1) included in each partial area Ai. M is the number of pixels included in the partial area Ai, and differs depending on the partial area Ai. The partial region feature amount calculation unit 124-x obtains a feature amount for each partial region Ai of the input pattern by, for example, a density pattern method. That is, the inner product value of the value Vij of the pixel Pij included in the partial area Ai and the weight value Wij determined in advance for the pixel Pij is calculated, and this result is obtained as the feature quantity Zi of the partial area Ai. . However, the present invention is not limited to this method, and other known methods such as a weighted direction histogram method may be used.

  By the way, for a group of similar characters including characters such as “large” and “thick”, the lower part of these similar characters is an important part for recognition. For such a group of similar characters, a partial area feature quantity calculation unit 124-different from the partial area feature quantity calculation unit 124-x provided for the similar character group including the character “B”. y is prepared. That is, the partial areas set in the partial area feature amount calculation unit 124-y are arranged densely in the lower part and sparsely in the other areas. As described above, the partial region feature amount calculation unit 124-x is provided for each group of similar characters having different parts important for recognition.

  The switching unit 125 selects one or more partial region feature value calculation units 124-x that provide an input pattern based on the recognition result of the primary recognition unit 110. For example, the switching unit 125 determines a group of similar characters including the character candidate that is the recognition result of the primary recognition unit 110, selects the partial region feature amount calculation unit 124-x corresponding to the group, and selects the partial region. An input pattern is given to the feature amount calculation unit 124-x. For this purpose, the pattern recognition apparatus 100 needs to hold a table (not shown) for knowing the correspondence between characters, groups, and partial region feature value calculation units 124-x. The switching unit 125 refers to this table to check the group to which the character candidate belongs, and further determines the partial region feature value calculation unit 124-x corresponding to this group. More specifically, for example, when two character candidates “bo” and “po” are obtained as a recognition result of the primary recognition unit 110, the switching unit 125 selects a group of similar characters including these characters. The partial area feature quantity calculation unit 124-x corresponding to the group, that is, the partial area feature quantity calculation unit 124-x in which the partial areas are densely arranged in the upper right part of the entire image area is selected, An input pattern is given to the partial region feature amount calculation unit 124-x.

  The feature vector generation unit 126 generates a feature vector representing the feature amount of the input pattern from the feature amount Zi for each partial region Ai obtained by the selected partial region feature amount calculation unit 124-x, and extracts the feature vector. The output from the unit 111 is given to the pattern identification unit 122. At this time, the product of the feature value Zi of the partial region Ai and the weight value Fi determined in advance for the partial region Ai is obtained, and this is used as the final feature value Di of the partial region Ai. A feature vector of the input pattern may be generated. Alternatively, a value obtained by converting the weight value Fi and the final feature amount Di by the nonlinear function G (x) may be used as the feature vector.

  The pattern identification unit 122 collates the feature vector output from the feature extraction unit 121 with the feature vector of the standard pattern stored in the dictionary 123, and outputs the character having the highest similarity as the final recognition result. .

  Here, one dictionary 123 and one pattern identification unit 122 are set. However, the dictionary and the pattern identification unit are divided and set in correspondence with the partial region feature value calculation unit 124-x, and switching is performed. A pattern may be configured to be identified using a dictionary and a pattern identification unit corresponding to the selection of the unit 125.

  Next, the operation of the pattern recognition apparatus 100 of this embodiment will be described.

  First, a pattern image of the target character is input to the primary recognition unit 110. The primary recognizing unit 110 extracts features from the input pattern by the feature extracting unit 111 to generate a feature vector, and gives it to the pattern identifying unit 112. The pattern identification unit 112 collates the input feature vector with the feature vector of the standard pattern stored in the dictionary 113 to determine one or more characters having a high degree of similarity, and uses these as candidate characters for recognition. Input to the next recognition unit 120.

  The secondary recognition unit 120 inputs the recognized character candidate given from the primary recognition unit 110 to the switching unit 125 of the feature extraction unit 121. The switching unit 125 checks the group to which the recognized character belongs, further determines the partial region feature quantity calculation unit 124-x corresponding to this group, and causes the partial region feature quantity calculation unit 124-x to recognize the pattern image of the recognition target character. give. At this time, if there are a plurality of recognized character candidates obtained from the primary recognition unit 110 and the groups to which the respective characters belong are different, detection is performed by separate partial region feature value calculation units 124-x corresponding to the respective groups. Give the input pattern of the target character.

  The partial region feature amount calculation unit 124-x calculates the feature amount for each partial region and provides the result to the feature vector generation unit 126. The feature vector generation unit 126 generates a feature vector representing the feature amount of the input pattern from the feature amounts for each partial region given by the partial region feature amount calculation unit 124-x, and gives the feature vector to the pattern identification unit 122. At this time, when feature values for each partial region are input from the plurality of partial region feature amount calculation units 124-x to the feature vector generation unit 126, corresponding feature vectors are generated and sent to the pattern identification unit 122. Given.

  When the pattern identification unit 122 acquires the feature vector, the pattern identification unit 122 compares the feature vector with the feature vector of the standard pattern stored in the dictionary 123 to calculate the similarity, and recognizes the character with the highest similarity. Output as.

  As described above, according to this embodiment, for each group of similar characters, it is possible to perform feature extraction with an increased focus on the parts important for the recognition of the similar characters. That is, by arranging the partial areas densely in the parts important for recognition in the pattern image and sparsely arranging the partial areas in the parts that are not so, the amount of information for calculating the feature amount of the parts important for recognition is increased. As a result, the identification accuracy of characters between similar characters can be increased.

(Second Embodiment)
Next, setting information of a plurality of partial regions in the partial region feature amount calculation unit 124-x is evaluated by evaluating a result obtained by recognizing a learning pattern each time changing one or a plurality of the partial regions. A method for generating the feature extraction parameter will be described.

  The following is a method for generating the feature extraction parameter of the partial region feature quantity calculation unit 124-x by a computer and a program that causes this to function as a learning tool.

  FIG. 4 is a flowchart showing a method of generating feature extraction parameters by this learning tool.

  First, the learning tool gives an initial learning pattern to the partial region feature amount calculation unit 124-x (step 401), and features of the respective partial regions Ai that are initially set in the partial region feature amount calculation unit 124-x. The amount Zi is calculated (step 402). Next, the learning tool multiplies the feature value Zi for each partial region Ai, which is the calculation result obtained from the partial region feature value calculation unit 124-x, by the weight value Fi set for each partial region, to thereby calculate each part. A feature amount Di of the area Ai is obtained (step 403).

  The learning tool generates a feature vector from the feature amount Di of each partial area Ai calculated in this way, and performs pattern recognition on the feature vector (step 404). The learning tool obtains the update amount of the weight value Fi of each partial area Ai based on the maximum similarity value obtained by pattern recognition (step 405), and the weight of each partial area Ai with these update amounts. The result of updating each value Fi is stored (step 406).

となる。ここでｄは正解カテゴリの相違度から不正解カテゴリの相違度を引いたもので、ｗは損失関数の微分である。αは学習重み値係数である。 As a method of calculating the update amount of the weight value Fi of each partial area Ai, for example, a probability drop method or the like can be cited, but other methods may be used. In the probability drop method, when pattern recognition is performed using the square of the Euclidean distance between the feature vector {Di} and the dictionary feature vector {Ei} as the degree of difference, the weight value Fi of each partial area Ai for one input pattern The renewal amount is

It becomes. Here, d is obtained by subtracting the difference degree of the incorrect answer category from the difference degree of the correct answer category, and w is a derivative of the loss function. α is a learning weight value coefficient.

  The learning tool repeats the above learning while changing the learning pattern, and repeatedly updates Fi. The process ends when the determined number of learnings is completed. When all patterns are used in this repetition, learning is repeated from the first pattern again (step 407).

  In the above description, the update of Fi is not immediately reflected in the feature amount calculation, provisional Fi ′ is prepared, Fi is substituted for Fi ′ as an initial value, and the update for all learning patterns is performed for provisional Fi ′. And Fi may be replaced with Fi ′ when the update for all patterns is completed. This series of learning is further performed a plurality of times.

(Third embodiment)
FIG. 5 is a flowchart showing another method of generating feature extraction parameters by the learning tool.

  First, the learning tool selects two partial areas from the partial area group that is initially set in the partial area feature amount calculation unit 124-x, and obtains a correlation value between these partial areas (step 501). Here, the square of the inner product of the vector consisting of the weight value of each pixel belonging to one partial area and the vector consisting of the weight value of each pixel belonging to the other partial area is divided by the square of the norm of each vector. Is the correlation value between the two partial areas. Since the two partial areas may be arranged so as to partially overlap each other, in this case, for the pixels that are commonly included in the two partial areas, the pixel weight values in the respective partial areas are used as they are. For the pixels that are included in only one partial area, the correlation value is calculated with the pixel weight value in the other partial area set to zero. The correlation value ranges from 0 to 1, and the higher the value, the higher the correlation, that is, the two partial areas are similar.

  The learning tool compares the correlation value with a predetermined threshold value (step 502), and determines that the correlation value is greater than the threshold value (YES in step 502), the two of the two partial regions Is deleted or two partial areas are replaced with these average partial areas (step 503).

  The average partial area is the area occupied by the two partial areas, and for the pixels included in the two partial areas in common, the average value of the weight values of the pixels in each partial area The weight value of the pixel in the new partial area is used. For pixels that are included only in one partial area, the weight value of that pixel in the other partial area is set to zero, and the average of the weight values of each pixel is This is the pixel weight value of the new partial area. If the average pixel weight value is less than a predetermined threshold value, this pixel may be excluded from the partial area.

  Based on the partial area feature amount calculation unit 124-x changed in this way, all or a part of the learning pattern is input to perform pattern recognition. That is, the feature amount for each partial area is calculated (steps 504 and 505). Create a dictionary using, for example, a certain percentage of the vectors obtained as the feature quantity calculation results, and then use, for example, a certain percentage of the feature vectors as an evaluation vector. Pattern recognition is performed by a method in which the character having the highest similarity and the value of the maximum similarity are obtained as recognition results by collating with the vector (step 506).

  Next, the learning tool determines whether or not the recognition accuracy has reached the acceptance standard (step 507). For example, it is evaluated whether the recognition accuracy is equal to or higher than a predetermined value. Alternatively, it is evaluated whether the performance improvement rate is equal to or less than a predetermined value. If the recognition accuracy does not reach the acceptance criteria, the learning tool starts a second learning cycle. The second learning cycle is performed by similarly obtaining a correlation value between two partial areas for the partial area group remaining in the first learning cycle. Thereafter, the learning tool again inputs the learning pattern to the partial region feature amount calculation unit 124-x, performs pattern recognition, and determines whether the recognition accuracy has reached the acceptance criterion. The above learning is repeated until a predetermined end condition is satisfied, for example, until the recognition accuracy reaches a pass criterion, or until a predetermined number of learning cycles are ended (step 508).

(Fourth embodiment)
FIG. 6 is a flowchart showing still another generation method of the feature extraction parameter by the learning tool.

  The learning tool selects a combination of K−i partial areas from among the K partial areas, and generates a feature vector from the feature amounts of these K−i partial areas (steps 602, 603, and 604). ). The initial value of i is 1 (step 601). Thus, a dictionary is created from part or all of the learning pattern, and pattern recognition is performed using this dictionary and part or all of the learning pattern as an evaluation pattern. Pattern recognition is performed by obtaining a character having the highest similarity and a value of the maximum similarity as a recognition result by collating the generated feature vector with the dictionary pattern (step 605).

  Next, the learning tool selects another combination of Ki (i = 1) partial areas by switching the partial areas that are not included in the generation of the feature vector (step 607), and performs the same processing. This is repeated until the processing for all combinations of the Ki partial areas is completed (step 606). Thereafter, the learning tool determines a combination of Ki (i = 1) partial areas that have the highest recognition accuracy in each recognition result (step 608).

  Next, the learning tool determines whether or not the recognition accuracy obtained by this determination has reached a pass criterion (step 609). If the recognition accuracy does not reach the acceptance standard, the learning tool further removes one partial area from the Ki (i = 1) partial areas determined in step 608. A combination of (i = 2) partial areas is selected (step 610), and the processing from step 602 to step 608 is similarly performed.

  The above process is repeated, and the combination of K−i partial areas when the recognition accuracy reaches the acceptance criterion is used as the final determination result. Even if the recognition accuracy does not reach the acceptance standard, this processing is also ended when the partial area Ai is narrowed down to a predetermined number (step 609).

(Fifth embodiment)
FIG. 7 is a flowchart showing still another generation method of the feature extraction parameter by the learning tool.

  First, the learning tool selects any two partial areas from the K partial areas that are initially set in the partial area feature amount calculation unit 124-x, creates these average partial areas, and creates a new one. It is set as a partial area and replaced with the original two partial areas (step 701). The method for creating the average partial area is as described in the third embodiment.

  Next, the learning tool inputs a learning pattern to the partial region feature amount calculation unit 124-x (step 702), and obtains a feature amount of each partial region (step 703). Subsequently, the learning tool generates a feature vector from the feature amount of each partial region (step 704). Thus, a dictionary is created from part or all of the learning pattern, and pattern recognition is performed using this dictionary and part or all of the learning pattern as an evaluation pattern. Pattern recognition is performed by obtaining a character having the highest similarity and a value of the maximum similarity as a recognition result by collating the generated feature vector with the dictionary pattern (step 705).

  Next, the learning tool returns the partial region of the partial region feature quantity calculation unit 124-x to the initial state, selects another two partial regions (step 707), returns to step 701, and averages these parts. The region is set as a new partial region, and the same processing as described above is performed (steps 702 to 705). When the above processing is completed for all partial regions (YES in step 706), the learning tool has the highest recognition accuracy among the recognition results obtained so far, for example, the one with the highest average similarity, When the correct answer of the learning pattern is known, the one with a high correct reading rate is selected, and the partial region set in the partial region feature amount calculation unit 124-x when the recognition result is obtained is reset (step) 708). As a result, the number of partial areas set in the partial area feature quantity calculation unit 124-x is K-1.

  Next, the learning tool determines whether or not the recognition accuracy obtained by this determination has reached a pass criterion (step 709). When the recognition accuracy does not reach the acceptance standard (NO in Step 709), the learning tool returns to Step 701 and creates an average partial area of two partial areas for K-1 partial areas. The same process is repeated.

  The above process is repeated, and a combination of partial areas when the recognition accuracy reaches the acceptance standard is set as a final determination result. Even if the recognition accuracy does not reach the acceptance standard, this processing is also ended when the partial area is narrowed down to a predetermined number.

  Note that when generating an average partial area, an average partial area of three or more partial areas may be created. Further, the average partial area may not be replaced with the original two partial areas, but may be added as another partial area on the original two partial areas.

(Sixth embodiment)
FIG. 8 is a flowchart showing still another generation method of the feature extraction parameter by the learning tool.

  First, the learning tool selects one partial area among the K partial areas that are initially set in the partial area feature amount calculation unit 124-x, and decomposes the partial area into a plurality of partial areas ( Step 801). For example, the pixels included in one partial area are divided into four groups of upper right, lower right, upper left, and lower left, and new partial areas are generated from the pixels belonging to each group. At this time, there may be pixels commonly included in a plurality of groups. The learning tool assigns a preset weight value to the new partial area. Next, the learning tool replaces the plurality of partial areas created in this way with the original partial areas set in the partial area feature quantity calculation unit 124-x.

  Next, the learning tool inputs a learning pattern to the partial region feature amount calculation unit 124-x (step 802), and obtains feature amounts of the partial regions (step 803). The learning tool generates a feature vector from the feature amount of each partial area (step 804), thereby creating a dictionary from a part or all of the learning pattern, and using this dictionary and part or all of the learning pattern as an evaluation pattern Pattern recognition. Pattern recognition is performed by obtaining a character having the highest similarity and a value of the maximum similarity as a recognition result by collating the feature vector with the dictionary pattern (step 805).

  Next, the learning tool returns the partial region of the partial region feature value calculation unit 124-x to the initial state, selects another partial region (step 807), returns to step 801, and repeats the same processing. . When the above processing is completed for all the partial regions (YES in Step 806), the learning tool has the highest recognition accuracy among the recognition results obtained so far, for example, the one with the highest average similarity, When the correct answer of the learning pattern is known, the one with a high correct reading rate is selected, and the partial region set in the partial region feature amount calculation unit 124-x when the recognition result is obtained is reset (step) 808).

  Next, the learning tool determines whether or not the recognition accuracy obtained by this determination has reached a pass criterion (step 809). If the recognition accuracy does not reach the acceptance standard (NO in step 809), the learning tool selects one of the reset partial areas, further decomposes, and repeats the same processing.

  The above process is repeated, and a combination of partial areas when the recognition accuracy reaches the acceptance standard is set as a final determination result. Even if the recognition accuracy does not reach the acceptance standard, this processing is also terminated when the rate of increase or increase in the number of partial areas or the number of decomposition layers reaches a predetermined value.

  Note that any number of partial areas may be decomposed at one time. Further, the disassembled partial area may not be replaced with the original partial area but may be added as another partial area on the original partial area.

(Seventh embodiment)
FIG. 9 is a flowchart showing still another generation method of the feature extraction parameter by the learning tool.

  First, the learning tool inputs a learning pattern to the partial region feature amount calculation unit 124-x in an initial state to obtain the feature amount of each partial region, and then generates a feature vector from the feature amount of each partial region. The character having the highest similarity and the value of the maximum similarity are obtained as a recognition result by matching the feature vector with the feature vector of the standard pattern of the detection target character type, and the recognition result is stored (step 901).

  Next, the learning tool adds one partial area to the K partial areas that are initially set in the partial area feature amount calculation unit 124-x (step 902). For example, in a region effective for recognition of similar characters corresponding to the partial region feature quantity calculation unit 124-x, a position where the partial regions are arranged relatively sparsely is detected, and a partial region having a certain size is added thereto. . As a method of detecting positions where the partial areas are relatively sparsely arranged, for each pixel in the area effective for recognition of similar characters corresponding to the partial area feature amount calculation unit 124-x, the partial area including the pixels is included. There is a method in which the number of the number is counted and a position where the number is smaller than a predetermined value is used as a detection result. Further, the weight values of the partial areas including the pixel may be integrated, and a position where the integrated value is lower than a predetermined value may be detected.

  Thereafter, the learning tool inputs the learning pattern again to the partial region feature amount calculation unit 124-x (step 903), and obtains the feature amount of each partial region (step 904). Subsequently, the learning tool generates a feature vector from the feature amount of each partial region (step 905), thereby creating a dictionary from part or all of the learning pattern, and evaluating this dictionary and part or all of the learning pattern. Pattern recognition is performed as a pattern for use. Pattern recognition is performed by obtaining the recognition result of the character having the highest similarity and the value of the maximum similarity by collating the generated feature vector with the dictionary pattern (step 906).

  Next, the learning tool compares the recognition result obtained before the addition of the partial region with the recognition result obtained after the addition (step 907), and if the improvement of the recognition accuracy is confirmed (YES in step 907), the addition is performed. The determined partial area is determined to be valid and is left as it is in the partial area feature amount calculation unit 124-x.

  Thereafter, the learning tool determines whether or not the recognition accuracy has reached a pass criterion (step 908). If the recognition accuracy does not reach the acceptance standard (NO in step 908), the learning tool returns to step 902 to add another partial region and perform the same processing. In addition, when the recognition accuracy reaches the acceptance standard (YES in step 908), or the increase rate or the increase number of the partial areas reaches a predetermined value even if the recognition accuracy does not reach the acceptance criterion. In addition, this process ends.

  If the improvement in recognition accuracy cannot be confirmed in the determination in step 907, the added partial area is deleted from the partial area feature amount calculation unit 124-x (step 909), and the process returns to step 902 to create a new one. Another partial area is added to and the same process is repeated.

(Eighth embodiment)
As shown in FIG. 10, a rectangular area formed by dividing the entire pattern image by a plurality of vertical lines Xi (i = 1,..., N) and horizontal lines Yi (i = 1,..., M) is defined as a partial area Ai. The feature amount for each partial area Ai may be calculated. These vertical lines Xi and horizontal lines Yi are set to have a pitch so that a rectangular area (partial area Ai) formed by these lines is dense at positions where recognition is effective, and sparse at other positions.

  As shown in FIG. 11, the vertical line Xi and the horizontal line Yi can be set so as to divide the pixel Pi. In this case, the pixel Pi is included in any of the adjacent partial areas Ai.

  A method of calculating the feature amount of the partial area Ai including the pixel Pi divided by the vertical line Xi and the horizontal line Yi in this way will be described below. The feature amount of the partial area Ai that is the calculation target is the sum of the feature amounts for each pixel Pi included in the partial area Ai that is the calculation target. As a method for obtaining a feature amount for each pixel Pi, there is a method by extracting features such as a density feature and a weighted direction histogram. In FIG. 11, only a part of the pixel Pi filled with diagonal lines is included in the partial area Ai to be calculated. Therefore, for these pixels Pi, only the value corresponding to the area ratio of the region belonging to the partial region Ai to be calculated is assigned to the feature amount of the partial region Ai to be calculated from the feature amount of the entire pixel.

  FIG. 12 is a flowchart illustrating a method for generating a partial region set in the partial region feature amount calculation unit 124-x of the present embodiment by a learning tool.

  The variable range and variable pitch of the x coordinate value of each vertical line Xi and the y coordinate value of each horizontal line Yi are determined, and the xy coordinate value of each vertical line Xi and each horizontal line Yi determined by this variable range and variable pitch. Keep all the combinations in the learning tool.

  First, the learning tool sets a plurality of partial areas Ai in the partial area feature amount calculation unit 124-x according to one of the combinations of the xy coordinate values of each vertical line Xi and each horizontal line Yi (step 1201), and the partial area features. A learning pattern is input to the quantity calculation unit 124-x (step 1202), and a feature quantity of each partial area Ai is obtained (step 1203). Next, the learning tool generates a feature vector from the feature amount of each partial area Ai (step 1204). Thus, a dictionary is created from part or all of the learning pattern, and pattern recognition is performed using this dictionary and part or all of the learning pattern as an evaluation pattern. Pattern recognition is performed by obtaining a character having the highest similarity and a value of the maximum similarity as a recognition result by collating the generated feature vector with the dictionary pattern (step 1205).

  Next, the learning tool returns to step 1201 and resets each partial area Ai of the partial area feature quantity calculation unit 124-x based on the next combination of the xy coordinate values of each vertical line Xi and each horizontal line Yi. The same processing as above is performed. This process is repeated for all combinations of the xy coordinate values of each vertical line Xi and each horizontal line Yi held in the learning tool (step 1206).

  After this, the learning tool has a high correct reading rate when it has the highest recognition accuracy among the recognition results obtained so far, for example, when the average similarity is high or when the correct answer of the learning pattern is known. When the recognition result is obtained, the partial area Ai set in the partial area feature amount calculation unit 124-x is reset (step 1207).

  Further, the vertical line Xi and the horizontal line Yi may be deleted or added, and learning may be performed by obtaining the recognition accuracy each time. At this time, the recognition accuracy may be obtained by changing the number of vertical and horizontal lines to be deleted or added.

(Ninth embodiment)
In the first embodiment, the partial region feature amount calculation unit 124-x is provided for each group of similar characters, but a plurality of partial region feature amount calculation units may be provided for each group of similar characters. For example, in FIG. 13, 124-01 and 124-02 are a plurality of partial region feature quantity calculation units provided for each group of similar characters, and 124-11 and 124-12 are for other groups of similar characters. Are a plurality of partial region feature amount calculation units provided in.

  As shown in the example of FIG. 14, for example, a character handwritten with “bo” has a difference in, for example, the position and shape of a muddy point depending on how to write. A plurality of partial region feature quantity calculation units 124-01 and 124-02 (124-11 and 124-12) for each group of similar characters lay out the layout of each partial region in accordance with such a difference in handwritten character features. Is set.

  Next, the operation of the pattern recognition apparatus 200 of this embodiment will be described.

  First, the input pattern of the detection target character is input to the primary recognition unit 110. In the primary recognition unit 110, features are extracted from the input pattern by the feature extraction unit 111 and a feature vector is given to the pattern identification unit 112. The pattern identification unit 112 collates the input feature vector with the feature vector of the standard pattern of the detection target character type stored in the dictionary 113 to determine one or more character candidates having a high degree of similarity. This character candidate is input to the secondary recognition unit 120 as a recognition result of the primary recognition unit 110.

  In the secondary recognition unit 120, the character candidate is given to the switching unit 125 in the feature extraction unit 121. The switching unit 125 checks the group to which the character candidate belongs, and further determines the partial region feature amount calculation unit 124-x corresponding to this group. For example, the switching unit 125 determines that the plurality of partial region feature value calculation units 124-01 and 124-02 correspond to the group, and recognizes these partial region feature value calculation units 124-01 and 124-02. Gives the input pattern of the target character. At this time, when there are a plurality of character candidates obtained from the primary recognition unit 110 and the groups to which the respective character candidates belong are different, the recognition target characters are sent to the partial region feature quantity calculation unit 124-x corresponding to each group. Are given respectively.

  Each of the partial region feature amount calculation units 124-01 and 124-02 calculates the feature amount for each partial region, and gives the result to the feature vector generation unit 126. In the feature vector generation unit 126, a feature vector representing the feature amount of the input pattern is generated from the feature amounts for each partial region given by the partial region feature amount calculation units 124-01 and 124-02. The result is given to the pattern identification unit 122 as an output of the feature extraction unit 121. The pattern identification unit 122 calculates the similarity by collating with the feature vector of the standard pattern stored in the dictionary 123 for each feature vector, and the character having the highest similarity for each feature vector is calculated. The value of the maximum similarity is output to the comprehensive determination unit 127. The overall determination unit 127 compares the two similarities and outputs the character with the higher recognition result as the final recognition result.

  Needless to say, three or more partial region feature value calculation units may be provided for each group of similar characters.

(Tenth embodiment)
The functions of the learning tool described so far may be implemented in the pattern recognition apparatus 100 as learning means.

  FIG. 15 is a block diagram illustrating a configuration of a pattern recognition apparatus 300 in which the learning unit 128 is mounted. The learning unit 128 performs learning for optimizing the partial region set in the partial region feature amount calculation unit 124-x by any of the methods described above, and the result is used as the partial region feature amount calculation unit 124. Reflect in -x. The learning pattern storage unit 129 stores a learning pattern that is an input pattern for learning.

  FIG. 16 is a flowchart regarding the learning operation of the pattern recognition apparatus 300 in which the learning unit 128 is mounted.

  The learning unit 128 acquires the recognition result obtained by the pattern identification unit 122 of the secondary recognition unit 120 (step 1601), and determines whether to perform learning based on the recognition result (step 1602). For example, when the maximum similarity value obtained as a recognition result is lower than a predetermined threshold value, the learning unit 128 determines to perform learning.

  The learning unit 128 first stores the input pattern when the maximum similarity value is less than the threshold value in the learning pattern storage unit 129 (step 1603). Thereafter, the learning unit 128 waits for the conditions necessary for performing learning to be satisfied, and starts processing for learning (step 1604). Here, the condition necessary for performing the learning is, for example, that the number of input patterns stored in the learning pattern storage unit 129 reaches a predetermined number necessary for the learning, and the state of the pattern recognition device 300 is the learning. This is a state suitable for performing, for example, when the pattern recognition apparatus 300 is not in operation.

  In order to start learning, the learning unit 128 reads the learning pattern stored in the learning pattern storage unit 129 and inputs it to the partial region feature amount calculation unit 124-x, and the feature vector generation unit 126, pattern identification unit Based on the recognition result obtained through 122, an optimum partial region is generated. The specific procedure is as described in the above embodiment.

とする。ｉ，ｊは、それぞれ縦位置、横位置である。一方、パターン画像上に複数のサンプル点を設定し、その位置を、

とする。この位置におけるフィルタを

とする。また、フィルタ後に得られた値をサプレスする関数を導入する。この関数にはたとえば、

が考えられるが、これに限定しなくともよい。 (Eleventh embodiment)
A learning tool modified and improved from the second embodiment will be described.
The pixel value of the pattern image

And i and j are a vertical position and a horizontal position, respectively. On the other hand, multiple sample points are set on the pattern image, and their positions are

And Filter at this position

And In addition, a function that suppresses the value obtained after filtering is introduced. For example, this function

However, the present invention is not limited to this.

とおいて、

で与えられる。ここで出力ベクトルの要素値は各位置ｓ_ａ，ｒ_ｂについて求めるため、次元はｎ×ｍとなる。 Feature extraction is when the element value of the output vector is χ _ab ,

Anyway,

Given in. Element value where the output vector each position s _a, to determine the r _b, dimension becomes n × m.

で定義する。 The degree of difference is determined by the feature vector χ = {χ _ab } obtained in this way and the dictionary vector φ = {φ _ab }.

Define in.

  The dictionary vector is determined for each character type, that is, for each category, and recognition is performed by outputting the category having the smallest difference as the correct category.

なので、

となり、

と書ける。ここでw（ｄ）は窓関数で、損失関数を微分したものである。損失関数とは、正解カテゴリの相違度Ｄ_okと不正解カテゴリの相違度Ｄ_errの差であるｄ＝Ｄ_ok−Ｄ_err上の関数で、右肩上がりの関数である。更新式におけるαaは学習係数で大きな値ほど学習強度が高く、学習時の辞書の変動幅が大きい。これはうまく学習が進むように適切な値を実験的に求めて適用する。更新式における上側の符号は正解カテゴリに対する更新、下側は不正解カテゴリに対する更新をあらわす。この更新は正解、不正解どちらの場合も第１位、すなわち最小相違度となるカテゴリを用いる。しかしながら、それ以外のカテゴリに対しても同様な処理を施すようにすることも可能である。 Next, a method for obtaining the dictionary vector φ and the filter F by learning in this recognition system will be described. Hereinafter, learning is performed based on the probability descent method. First, D is differentiated using φ _ab as a parameter.

So,

And

Can be written. Here, w (d) is a window function, which is obtained by differentiating the loss function. The loss function is a function on d = D _ok −D _err , which is the difference between the difference category D _ok of the correct answer category and the difference degree D _err of the incorrect answer category, and is a function that rises to the right. Αa in the update formula has a higher learning intensity as the learning coefficient is larger, and the fluctuation range of the dictionary during learning is larger. In this method, an appropriate value is experimentally obtained and applied so that learning progresses well. The upper code in the update formula represents an update for the correct answer category, and the lower code represents an update for the incorrect answer category. This update uses the first rank, that is, the category with the smallest difference in both correct and incorrect answers. However, it is possible to perform the same processing for other categories.

これによりフィルタの更新式は

となる。βは学習係数。 Next, D is differentiated using F _ab (k, h) as a parameter.

This makes the filter update formula

It becomes. β is the learning coefficient.

  Based on the result of character recognition and this update formula, the dictionary and the filter are slightly changed and updated to perform character recognition. Learning is performed by repeating this update.

  Based on the recognition accuracy, the learning is terminated according to a preset termination condition. By this learning method, a dictionary and a filter with higher performance can be created.

  In the present invention, in the feature extraction process of the above embodiment, ρ (t) may be removed to form a linear filter. In the recognition process using the filter created in this way, recognition by another recognition system may be used instead of the Euclidean distance. Of course, the recognition system used in the learning is not limited to the Euclidean distance, and anything that can be applied to the probability descent method, such as simple similarity using the inner product, may be used.

  In the above-described embodiment, the filter is described as one type. However, a plurality of types of filters may be created. In this case, all or some of the feature values output from the plurality of filters are collected as the feature vector input to the pattern identification unit. The update formula in the learning method may be the same as that in the above embodiment. In learning, learning may be performed with one type of filter at first, and a new filter may be added when learning has progressed to some extent, and thereafter, learning may be performed with the two filters. The same applies to the third and subsequent filters.

  In the above description, the pixel value is used for description. However, this may be a feature value obtained using each pixel and its surrounding pixels.

  As mentioned above, although embodiment which employ | adopted this invention for character recognition was described, this invention is applicable to recognition of the object which can recognize patterns, such as an image and a mark.

In the above embodiment, the case where the present invention is applied to the secondary recognition unit that narrows down the final recognition result from the characters narrowed down by the primary recognition unit has been described. However, primary recognition is not performed without adopting such a configuration. It can also be applied to the case where a final recognition result is obtained only by the above. It is also possible to carry out learning by combining the respective means described in the above embodiment.
It should be noted that the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the scope of the present invention.

1 is a block diagram showing a configuration of a pattern recognition apparatus according to a first embodiment of the present invention. The figure which shows the example of the input pattern image. The figure which shows the example of the several partial area | region which was set numerically nonuniformly in the whole area | region of a pattern image. The flowchart which shows the production | generation method of the feature extraction parameter by a learning tool. The flowchart which shows another production | generation method of the feature extraction parameter by a learning tool. The flowchart which shows another production | generation method of the feature extraction parameter by a learning tool. The flowchart which shows another production | generation method of the feature extraction parameter by a learning tool. The flowchart which shows another production | generation method of the feature extraction parameter by a learning tool. The flowchart which shows another production | generation method of the feature extraction parameter by a learning tool. The figure which shows another method of producing | generating several partial area | regions in the whole area | region of a pattern image FIG. 11 is a diagram for explaining a method for calculating a feature amount of a partial region including pixels divided by a vertical line and a horizontal line in the partial region generation method of FIG. 10. FIG. 11 is a flowchart illustrating a method for generating a partial region set in the partial region feature amount calculation unit by a learning tool in the partial region generation method of FIG. The block diagram which shows the structure of the pattern recognition apparatus which provided the some partial area | region feature-value calculation part for every group of similar characters. The figure which shows the partial area | region set to the some partial area | region feature-value calculation part provided for every similar character group of FIG. It is a block diagram which shows the structure of the pattern recognition apparatus which mounted the learning part. The flowchart regarding the learning operation | movement of the pattern recognition apparatus which mounted the learning part of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Pattern recognition apparatus 121 ... Feature extraction part 122 ... Pattern identification part 123 ... Dictionary 124 ... Partial region feature-value calculation part 125 ... Switching part 126 ... Feature vector generation part 127 ... Comprehensive determination part 128 ... Learning part 129 ... Learning pattern storage Part

Claims (2)

A partial region feature amount calculation unit that calculates a feature amount of each of the plurality of partial regions set in the image region with respect to the input pattern image;
A feature vector generation unit that generates a feature vector of the input pattern from the feature amounts of each of the plurality of partial regions calculated by the partial region feature amount calculation unit;
A pattern identification unit that identifies the input pattern using the feature vector output from the feature vector generation unit;
The feature vector generation unit includes:
In the partial region feature amount calculation unit, the pattern value inputted from the feature amount corresponding to the partial region calculated by the product value of the pixel value in the partial region or the feature value corresponding to the pixel and the weight corresponding to the pixel is calculated. Generating a feature vector,
The weight value corresponding to the pixel for each partial region is
The loss function for obtaining the loss value of the degree of difference or similarity with the standard pattern obtained by recognizing the learning pattern is differentiated by the variable representing the weight value, and the differentiated value and the learning pattern are recognized. A pattern recognition apparatus that has been changed by evaluating a result obtained by performing the above. A partial region feature amount calculation unit that calculates a feature amount of each of the plurality of partial regions set in the image region with respect to the input pattern image;
A feature vector generation unit that generates a feature vector of the input pattern from the feature amounts of each of the plurality of partial regions calculated by the partial region feature amount calculation unit;
A pattern identification unit that identifies the input pattern using the feature vector output from the feature vector generation unit;
The feature vector generation unit includes:
Generates a feature vector of the input pattern from the feature value obtained by multiplying the feature value of each of the plurality of partial regions calculated by the partial region feature value calculation unit by the weight value determined for each partial region. To do,
The weight value for each partial area is
The loss function for obtaining the loss value of the degree of difference or similarity with the standard pattern obtained by recognizing the learning pattern is differentiated by the variable representing the weight value, and the differentiated value and the learning pattern are recognized. A pattern recognition apparatus that has been changed by evaluating a result obtained by performing the above.

Images