JP2007018176A - Learning device, learning method, learning program, recording medium, and device and method for pattern recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a learning device, a learning method, a learning program, and a recording medium, enabling more accurate learning, and a device and a method for pattern recognition enabling more accurate pattern recognition. <P>SOLUTION: For a value deviating greater from a region in which values are most densely distributed, a weight for use when calculating a degree of scatter in a class is varied greater. The degree of scatter is calculated on the basis of a value corresponding to a learning pattern being classified by a discrimination means learned by feedbacking the calculated estimation value. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a learning device, a learning method, a learning program, a recording medium, a pattern recognition device, and a pattern recognition method, and in particular, a learning device, a learning method, a learning program, a recording medium, a pattern recognition device that perform processing based on learning, and The present invention relates to a pattern recognition method.

  2. Description of the Related Art In recent years, application fields of image processing have been expanding due to lower prices of image capturing devices such as CCD (Charge Coupled Device) and image processing equipment such as personal computers. In particular, the application of image processing to defect inspection in the industrial field is extensive. For example, image processing is becoming more and more necessary in order to improve the yield and quality of product devices and to reduce costs by reducing the number of visual inspectors.

  Therefore, at present, various image processing algorithms have been developed for defect inspection in accordance with target products and devices. This is because the image processing algorithm for defect inspection is generally designed exclusively for each defect target, and the image processing algorithm once designed cannot be applied to other defect inspections in many cases. In other words, there is almost no reusability of the image processing algorithm once developed, and the development man-hours increase in proportion to the expansion of application fields and inspection objects.

  Development of an image processing algorithm consists of a design process and an adjustment process. Especially in the adjustment process, various parameter changes and evaluations are repeated by trial and error. The development flow of the image processing algorithm is roughly as follows.

  First, a plurality of defect image data samples including a target defect and a plurality of non-defective image data samples including no defect are acquired. Then, a limit defect image data sample that is a defect image data sample including a defect that is considered difficult to detect is selected from the acquired defect image data samples. Thereafter, a rough procedure of basic image processing, such as image normalization, noise removal, binarization processing, and binary image shape measurement, is determined from the features of the defect and the background. The process up to here is positioned as a design work in the development of an image processing algorithm.

  Next, the number of filter processes, filter shapes, threshold values, etc. that constitute the image processing procedure so that only defects are detected when the above basic image processing is applied to the limit defect image data sample. Change the parameter value. Then, by evaluating the limit defect image data sample to which the image processing after changing the parameter value is applied, a base image processing algorithm is created. Here, the evaluation is best when only a target defect is detected and nothing else is detected.

  Further, the defect included in all the defect image data samples as well as the limit defect image data sample is detected, and the parameter change and evaluation of the base image processing algorithm are repeated until nothing is detected from the non-defective image data sample. This completes the development of the image processing algorithm.

  However, it is usually difficult to develop an image processing algorithm that detects defects contained in all defect image data samples and detects nothing from all non-defective image data samples. That is, a case where some of the true defects cannot be detected (undetected) or a case where a pseudo defect which is a non-defect portion is detected (overdetection) is inevitable. Hereinafter, a defect to be detected by the image processing algorithm is referred to as a true defect. In addition, a defect that is erroneously determined as a defect even though it should not be detected by the image processing algorithm is called a pseudo defect.

  Hereinafter, image data for generating an image processing algorithm is also referred to as learning image data. Here, the learning image data is image data in which data such as position, size, shape, etc. in an image of an extraction target object such as a defect (hereinafter referred to as defect sample information) is given in advance by an operator.

  The operation of designating defect sample information in the image of the extraction object such as a defect is still performed by the operator. Accordingly, there is a case where the designation of the defect sample information varies depending on the operator's sense, experience, fatigue state, or the like, the defective part is missed, or a non-defective part is erroneously designated as a true defect.

  Hereinafter, these are collectively referred to as defect sample information fluctuation. Here, the variation in the designation of the defect sample information is, for example, a case where the position of the defect is designated by being shifted, or a case where the designation of the shape of the defect is too large or too small.

  Such a variation in defect sample information affects the feature amount extracted by the feature extraction unit, and further greatly affects the evaluation directly calculated by the evaluation unit. Eventually, it becomes difficult to correctly learn the image processing algorithm.

  In the image processing algorithm, over-detection occurs when parameters are adjusted to increase the defect detection sensitivity of the image processing algorithm in order to avoid undetection. In the image processing algorithm, if the defect detection sensitivity is lowered in order to avoid overdetection, non-detection occurs. That is, there is a trade-off between undetected and overdetected. Therefore, parameter adjustment and evaluation work in image processing algorithm development is very difficult. As described above, in the development of an image processing algorithm, a great amount of man-hours are required for these empirical and trial and error parameter adjustments and evaluation operations.

So, Zhou Muei, 4 others, “Improvement of 3D image processing expert system 3D-IMPRESS-Pro and application to automatic configuration of lung cancer shadow detection procedure”, Journal of Computer Aided Diagnosis Society, May 2000, Vol. In No. 1 (Non-Patent Document 1), a technique for automatically generating an image processing algorithm in consideration of a case where detection is performed without missing an abnormal shadow and a case where a part not corresponding to a true abnormal shadow is detected (hereinafter, conventional) Technology A) is also proposed. Specifically, an evaluation value considering a predetermined non-detection rate and a predetermined overdetection rate is calculated, and an image processing algorithm is learned based on the evaluation value.
Ei Zhou Mukai, 4 others, “Improvement of 3D image processing expert system 3D-IMPRESS-Pro and application to automatic configuration of lung cancer shadow detection procedure”, Journal of Computer Aided Imaging Society, May 2000, Vol. 1

  However, in the conventional technique A, the evaluation value is calculated by paying attention only to the number of learning patterns that are not detected or overdetected. In other words, the evaluation value does not consider what value the feature amount of the learning pattern is. Therefore, it is impossible to determine the superiority or inferiority of two learning algorithms having the same undetected rate and overdetected rate.

  As described above, the variation of the feature amount is caused by the variation of the defect sample information. However, no learning method has been proposed so far in consideration of various defect sample information fluctuations.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a learning apparatus that can learn more accurately even when there is a variation in defect sample information.

  Another object of the present invention is to provide a learning method that allows more accurate learning even when there is a variation in defect sample information.

  Still another object of the present invention is to provide a learning program that can learn more accurately even when there is a variation in defect sample information.

  Still another object of the present invention is to provide a recording medium that records a learning program that can be more accurately learned even when there is a variation in defect sample information.

  Still another object of the present invention is to provide a pattern recognition apparatus that can more accurately identify a pattern.

  Still another object of the present invention is to provide a pattern recognition method capable of more accurately identifying a pattern.

  In order to solve the above-described problem, a learning device according to an aspect of the present invention includes a pre-processing unit that pre-processes a plurality of learning patterns in which an assigned class is designated in advance, and a plurality of learning patterns that have been subjected to pre-processing. A feature extraction means for extracting a feature quantity from each of the above, an identification means for classifying each of the plurality of learning patterns into one of a plurality of classes based on the value of the corresponding feature quantity, and a plurality of classes For each feature class value corresponding to each of a plurality of learning patterns, a class dispersion degree indicating a distribution state in the corresponding class is calculated for each of a plurality of types of classes. A calculation means for calculating the degree of inter-class dispersion indicating the distribution state between classes, and the evaluation value increases as the degree of intra-class dispersion decreases, and the degree of evaluation increases as the degree of inter-class dispersion increases. And an evaluation means for calculating an evaluation value that satisfies one or both of the increase of the evaluation value, and the evaluation means feeds back the calculated evaluation value so that the evaluation value becomes high, so that the preprocessing means , Learning at least one of the feature extraction means or the identification means, and the calculation means increases the degree of dispersion in the class as the feature value that is far from the range in which the feature value for each of a plurality of types of classes is distributed most frequently. The weight for calculation is changed.

  Preferably, the calculation unit reduces the weights as the feature amount values away from the range in which the feature amount values for each of the plurality of types of classes are most distributed, and the plurality of types corresponding to each of the plurality of types of classes. Intra-class average based on multiple feature values corresponding to each of the learning patterns is calculated for each of multiple types of classes, and the intra-class dispersion level indicating the degree of dispersion from the in-class average is calculated for each of multiple types of classes. And the inter-class scatter degree is calculated based on the average within the class corresponding to each of a plurality of types of classes.

  Preferably, the calculating means calculates a plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of a plurality of types of classes and an in-class average candidate corresponding to the corresponding class, Calculate for each class of multiple types, and for each of the multiple distances, calculate the weighted distance for each of the multiple types of classes. Then, by averaging the calculated multiple weighted distances based on the number of learning patterns corresponding to each of the multiple types of classes, the within-class dispersion degree candidates are calculated for each of the multiple types of classes. The intra-class average candidate that minimizes the dispersion degree candidate is calculated as the intra-class average, and the intra-class dispersion degree candidate that is minimized is calculated as the intra-class dispersion degree candidate.

  Preferably, the averaging for calculating the degree of dispersion within the class by the calculating means is an arithmetic average.

  Preferably, a plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of a plurality of types of classes and an in-class average corresponding to the corresponding class are set for each of the plurality of types of classes. For each of a plurality of distances, a power distance that is a power of a predetermined power value is calculated for each of a plurality of types of classes, and a median value based on the distribution of the calculated power distances is defined as the degree of dispersion within the class. Calculate for each of multiple types of classes.

Preferably, the power value is 2.
Preferably, the distance is an Euclidean distance.

  Preferably, the evaluation unit calculates, as an evaluation value, a value obtained by dividing the inter-class dispersion degree by the total value of the plurality of intra-class dispersion degrees corresponding to each of the plurality of types of classes.

  Preferably, the identification unit further includes an erroneous specification determination unit that determines that the learning pattern is classified into a class other than the previously specified belonging class as an erroneous specification, and the erroneous specification determination unit is a target of erroneous specification. The learned pattern is excluded from the learning pattern that is the target for calculating the evaluation value calculated by the evaluation means.

  Preferably, the learning pattern is an image, the preprocessing is processing for applying image processing based on a predetermined image processing algorithm to the learning pattern, and the evaluation unit changes the image processing algorithm based on the evaluation value.

  Preferably, the evaluation unit changes the image processing algorithm based on a genetic algorithm.

  A learning method according to another aspect of the present invention includes a step of performing preprocessing on a plurality of learning patterns for which an assigned class is specified in advance, and a step of extracting feature amounts from each of the plurality of learning patterns on which preprocessing has been performed. And classifying each of the plurality of learning patterns into one of a plurality of types of classes based on the corresponding feature value, and each of the plurality of learning patterns classified into each of the plurality of types of classes For the feature value corresponding to, calculate the degree of intra-class dispersion indicating the distribution state in the corresponding class for each of multiple types of classes, and further, the degree of inter-class dispersion indicating the distribution state between multiple types of classes The evaluation value increases as the degree of dispersion within the class decreases, and the evaluation value increases as the degree of dispersion between classes increases. A step of calculating an evaluation value, and the step of calculating the evaluation value includes at least a step of performing a pre-processing by feeding back the calculated evaluation value so as to increase the evaluation value, a step of extracting or a step of classifying The step of calculating includes the step of learning one, and the step of calculating is weighting when calculating the degree of dispersion within the class as the value of the feature amount that is far from the range in which the value of the feature amount for each of a plurality of types of classes is the most distributed The process of changing is included.

  Preferably, the step of calculating the degree of inter-class dispersion is performed by reducing the weights of feature values that are farther away from the range in which the feature values for each of the plurality of types of classes are the most distributed. A step of calculating an average within a class for each of a plurality of types of classes based on a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of the patterns, and an intra-class dispersion indicating a degree of dispersion from the intra-class average A step of calculating the degree for each of a plurality of types of classes, and a step of calculating a degree of scatter among classes based on an average within the class corresponding to each of the plurality of types of classes.

  Preferably, the step of calculating the inter-class dispersion degree includes a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of a plurality of types of classes, and an intra-class average candidate corresponding to the corresponding class. Calculating a plurality of distances for each of a plurality of types of classes, and for each of the plurality of distances, a weighted distance that is weighted so that the calculated value decreases as the value of each of the plurality of distances increases. The process of calculating for each of a plurality of types of classes, and the plurality of calculated weighted distances are averaged based on the number of learning patterns corresponding to each of the plurality of types of classes, thereby generating a plurality of in-class dispersion degree candidates. Calculate for each type of class, calculate the within-class average candidate that minimizes the within-class dispersion degree candidate as the within-class average, and select the within-class dispersion degree candidate that has the smallest within-class dispersion And a step of calculating a burr of candidates.

  Preferably, the averaging for calculating the inter-class dispersion degree in the step of calculating the inter-class dispersion degree is an arithmetic average.

  Preferably, the step of calculating the inter-class dispersion degree includes a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of a plurality of types of classes, and an in-class average corresponding to the corresponding class. A step of calculating a plurality of distances for each of a plurality of types of classes, a step of calculating, for each of the plurality of types of classes, a power distance that should be a power of a predetermined power value for each of the plurality of distances, And calculating a median value based on the power-distance distribution as the degree of dispersion within the class for each of a plurality of types of classes.

Preferably, the power value is 2.
Preferably, the distance is an Euclidean distance.

  Preferably, the step of calculating the evaluation value includes a step of calculating, as the evaluation value, a value obtained by dividing the inter-class dispersion degree by the total value of the plurality of intra-class dispersion degrees corresponding to each of the plurality of types of classes.

Preferably, the step of classifying further includes the step of determining that the learning pattern is classified into a class other than a previously designated affiliation class as an erroneous designation,
The step of determining that the designation is incorrect includes a step of excluding the learning pattern that is the target of erroneous designation from the learning pattern that is the target of calculating the evaluation value calculated by the step of calculating the evaluation value.

  Preferably, the learning pattern is an image, and the preprocessing is processing for performing image processing based on a predetermined image processing algorithm on the learning pattern, and the step of calculating the evaluation value is based on the evaluation value. The process of changing is included.

  Preferably, the step of calculating the evaluation value includes a step of changing the image processing algorithm based on a genetic algorithm.

  According to still another aspect of the present invention, a learning program for causing a computer to execute a learning process includes a step of performing preprocessing on a plurality of learning patterns in which a belonging class is designated in advance, and a plurality of preprocessed plurality of learning patterns. Extracting a feature amount from each of the learning patterns, classifying each of the plurality of learning patterns into one of a plurality of classes based on the corresponding feature value, and a plurality of classes For each feature value corresponding to each of a plurality of learning patterns classified into each class, the degree of intra-class dispersion indicating the distribution state in the corresponding class is calculated for each of a plurality of types of classes. A step of calculating the inter-class dispersion degree indicating the distribution state between the classes, the evaluation value increases as the intra-class dispersion degree decreases, and the inter-class dispersion degree. Calculating the evaluation value that satisfies one or both of the fact that the evaluation value increases as the degree increases, and the step of calculating the evaluation value causes the evaluation value to increase. Including a step of performing preprocessing by feeding back the calculated evaluation value, and a step of learning at least one of a step of extracting or classifying, wherein the step of calculating has the largest feature value for each of a plurality of types of classes. The step of changing the weight at the time of calculating the degree of dispersion in the class is included as the value of the feature amount that is far from the range in which many are distributed.

  According to still another aspect of the present invention, the recording medium is a medium on which a learning program is recorded.

  A pattern recognition apparatus according to still another aspect of the present invention recognizes a pattern of a processing target image using an image processing algorithm.

  A pattern recognition method according to still another aspect of the present invention recognizes a pattern of a processing target image using an image processing algorithm.

  The learning device according to the present invention calculates an evaluation value that increases as the degree of dispersion within the class decreases, and calculates the degree of dispersion within the class as the value is farther from the range in which the values are most distributed. Change the weight of. The degree of dispersion within the class is calculated based on a value corresponding to the learning pattern classified by the discriminating means that learns by feeding back the calculated evaluation value.

  Therefore, a more accurate evaluation value can be calculated because a more accurate dispersion degree within the class can be calculated. As a result, it is possible to learn more accurately.

  The learning method according to the present invention calculates an evaluation value that increases as the degree of dispersion within the class decreases. When calculating the degree of dispersion within a class as a value that is far from the range in which the values are most distributed. Change the weight of. The degree of dispersion within the class is calculated based on the value corresponding to the learning pattern classified by the classification step of learning by feeding back the calculated evaluation value.

  Therefore, a more accurate evaluation value can be calculated because a more accurate dispersion degree within the class can be calculated. As a result, it is possible to learn more accurately.

  The learning program according to the present invention calculates an evaluation value that increases as the degree of dispersion within the class decreases. When calculating the degree of dispersion within the class, the value farther from the range in which the values are most distributed is calculated. Change the weight of. The degree of dispersion within the class is calculated based on a value corresponding to the learning pattern classified by the classification step of learning by feeding back the calculated evaluation value.

  Therefore, a more accurate evaluation value can be calculated because a more accurate dispersion degree within the class can be calculated. As a result, it is possible to learn more accurately.

  The recording medium according to the present invention records a learning program. The learning program calculates an evaluation value that increases as the degree of dispersion within the class decreases, and changes the weight when calculating the degree of dispersion within the class as the value moves away from the most widely distributed range. Let The degree of dispersion within the class is calculated based on a value corresponding to the learning pattern classified by the classification step of learning by feeding back the calculated evaluation value.

  Therefore, a more accurate evaluation value can be calculated because a more accurate dispersion degree within the class can be calculated. As a result, it is possible to learn more accurately.

  The pattern recognition apparatus according to the present invention recognizes a pattern of a processing target image using an image processing algorithm for further increasing the identification rate.

Therefore, there is an effect that the pattern can be identified more accurately.
The pattern recognition method according to the present invention recognizes a pattern of an image to be processed using an image processing algorithm for increasing the identification rate.

  Therefore, there is an effect that the pattern can be identified more accurately.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated.

<First Embodiment>
FIG. 1 is a block diagram showing an internal configuration of learning apparatus 500 in the present embodiment. The learning device 500 is, for example, a computer such as a PC (Personal Computer) or a workstation. Note that FIG. 1 also shows a recording medium 555 and an image data input unit 600 for explanation. A recording medium 555 stores a program 180 described later. That is, the program 180 is recorded on a medium or the like and distributed as a program product. The recording medium 555 is also distributed as a program product.

  Although details will be described later, the image data input unit 600 has a function of transmitting image data to the learning device 500.

  With reference to FIG. 1, a display unit 530, a mouse 542, and a keyboard 544 are connected to the learning apparatus 500.

  The display unit 530 has a function of displaying various types of information as characters, images, and the like to the user. Display unit 530 displays an image based on the image data output from learning device 500. The display unit 530 is a liquid crystal display (LCD (Liquid Crystal Display)), a CRT (Cathode Ray Tube), an FED (Field Emission Display), a PDP (Plasma Display Panel), an organic EL display (Organic Electroluminescence Display), a dot matrix, or the like. Any other display device of an image display method may be used.

  The mouse 542 is an interface for the user to operate the learning device 500. The keyboard 544 is an interface for the user to operate the learning apparatus 500. Note that the display unit 530 and the keyboard 544 may be provided inside the learning apparatus 500. Moreover, the structure provided in the inside of the learning apparatus 500 may be sufficient as the mouse | mouth 542 as a touchpad etc., for example. Further, a pen tablet may be connected to the learning apparatus 500 as an input device.

  The learning device 500 includes a control unit 510, a data temporary storage unit 522, a storage unit 520, a communication unit 560, a communication unit 562, a VDP (Video Display Processor) 532, and a CGROM (Character Graphic Read Only Memory) 534. A VRAM (Video Random Access Memory) 536, an input unit 540, and a recording medium access unit 550.

  The CGROM 534 stores image data such as font data and graphic data for the VDP 532 to generate an image displayed on the display unit 530.

  The storage unit 520 stores a program 180 for causing the control unit 510 to perform processing described later, a learning image database 182 for causing the image processing algorithm to learn, various other programs, data, and the like. The storage unit 520 is accessed by the control unit 510. Note that the learning image database 182 is a database composed of data of a plurality of types of learning images that are different from each other. Details of the learning image will be described later.

  The storage unit 520 is a hard disk capable of storing a large amount of data. Note that the storage unit 520 is not limited to a hard disk, and may be any medium (for example, a flash memory) that can hold data in a nonvolatile manner even when power is not supplied.

  Control unit 510 has a function of performing various types of processing, arithmetic processing, and the like for each device inside learning device 500 in accordance with program 180 stored in storage unit 520. The control unit 510 includes a microprocessor, an FPGA (Field Programmable Gate Array), which is an LSI (Large Scale Integration) that can be programmed, and an ASIC (Application, which is an integrated circuit designed and manufactured for a specific application. Specific Integrated Circuit) or any other circuit having an arithmetic function may be used.

  Control unit 510 also instructs VDP 532 to generate an image and display the image on display unit 530 in accordance with program 180 stored in storage unit 520 (hereinafter, also referred to as “drawing instruction”). It's out.

  The VDP 532 is connected to the display unit 530. The VDP 532 reads necessary image data from the CGROM 534 in accordance with a drawing instruction from the control unit 510 and generates an image using the VRAM 536. The VDP 532 reads out the image data stored in the VRAM 536 and causes the display unit 530 to display an image based on the image data.

The VRAM 536 has a function of temporarily storing an image generated by the VDP 532.
The data temporary storage unit 522 is accessed as data by the control unit 510 and used as a work memory for temporarily storing data.

  The data temporary storage unit 522 is a high-speed RAM (Random Access Memory), SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), SDRAM (Synchronous DRAM), or double data rate mode capable of temporarily storing data. DDR-SDRAM (Double Data Rate SDRAM), which is an SDRAM with excellent data transfer function, RDRAM (Rambus Dynamic Random Access Memory), DRAM that adopts high-speed interface technology developed by Rambus, Direct-RDRAM (Direct Rambus Dynamic) Random Access Memory) or any other circuit having a configuration capable of storing and storing data in a volatile manner.

  A mouse 542 and a keyboard 544 are connected to the input unit 540. The user gives an instruction to the learning apparatus 500 using the mouse 542 or the keyboard 544. An input instruction from the mouse 542 or the keyboard 544 is transmitted to the control unit 510 via the input unit 540. Control unit 510 performs predetermined processing based on an input instruction from input unit 540.

  The recording medium access unit 550 has a function of reading the program 180 from the recording medium 555 on which the program 180 is recorded. The program 180 stored in the recording medium 555 is read from the recording medium access unit 550 and stored in the storage unit 520 by the operation (installation process) of the control unit 510.

  The installation processing program is stored in advance in the storage unit 520, and the installation processing is performed by the control unit 510 based on the installation processing program.

  Note that the program 180 may not be installed in the storage unit 520. In this case, control unit 510 reads program 180 stored in recording medium 555 via recording medium access unit 550, and performs predetermined processing based on program 180.

  The recording medium 555 includes a DVD-ROM (Digital Versatile Disk Read Only Memory), a DVD-R (Digital Versatile Disk Recordable), a DVD-RAM (Digital Versatile Disk Random Access Memory), a DVD + RW (Digital Versatile Disk Re-Writable), and a DVD. -RW, CD-ROM (Compact Disk Read Only Memory), MO (Magneto Optical Disk), floppy (registered trademark) disk, CF (Compact Flash) card, SM (Smart Media (registered trademark)), MMC (Multi Media Card) ), SD (Secure Digital) memory card, Memory Stick (registered trademark), xD picture card, USB memory, magnetic tape, and other non-volatile memory.

  Communication unit 560 exchanges data with control unit 510. Communication unit 560 exchanges data with image data input unit 600 in a wired or wireless manner.

  Control unit 510 transmits control data for controlling image data input unit 600 via communication unit 560. Control unit 510 receives data corresponding to the transmitted control data from image data input unit 600 via communication unit 560.

  The communication unit 560 may be any of USB (Universal Serial Bus) 1.1, USB 2.0, and other communication interfaces for performing serial transfer. The communication unit 560 may be a communication interface using Ethernet (registered trademark). The communication unit 560 may be any of a Centronics specification, IEEE 1284 (Institute of Electrical and Electronic Engineers 1284), and other communication interfaces that perform parallel transfer. The communication unit 560 may be any one of communication interfaces using IEEE 1394 or other SCSI standards.

  The communication unit 560 may be any of wireless LAN standards IEEE802.11a, IEEE802.11b, IEEE802.11g, and other communication interfaces for performing data communication using a wireless technology.

  Communication unit 562 exchanges data with control unit 510. Since communication unit 562 is similar to communication unit 560, detailed description will not be repeated.

  Communication unit 610 exchanges data with communication unit 562 and network 620. The network 620 is an external network such as the Internet. The communication unit 610 is a communication interface (for example, a router) using Ethernet (registered trademark). The communication unit 610 and the network 620 exchange data wirelessly or by wire. The wire is, for example, a telephone line or an optical fiber cable.

  Therefore, learning apparatus 500 in the present embodiment can also download a program from network 620 via communication unit 610 and communication unit 562, and store the program in storage unit 520. In this case, the downloaded program is the program 180.

  Control unit 510 performs predetermined processing according to a program (program 180) downloaded from network 620. The download program is stored in advance in the storage unit 520, and the download process is performed by the control unit 510 based on the download program.

The learning apparatus 500 having the above configuration generates an image processing algorithm.
FIG. 2 is a block diagram showing the internal configuration of the image data input unit 600. Note that FIG. 2 also shows a communication unit 560 in the learning apparatus 500 for the sake of explanation.

  Referring to FIG. 2, image data input unit 600 has a function of irradiating light to inspection object 650 to be inspected installed on stage 640 and converting reflected light of inspection object 650 into data. The inspection object 650 may be a liquid crystal panel, a semiconductor, an electronic component, plastic, metal, wood, paper, cloth, or the like.

  The image data input unit 600 includes a control unit 610, a light source 620, and a camera 630. Control unit 610 exchanges data with communication unit 560 in learning device 500. Control unit 610 has the same configuration and function as control unit 510 described above. Control unit 610 controls light source 620 and camera 630 based on control data received from control unit 510 via communication unit 560.

The light source 620 applies light to the inspection object 650 under the control of the control unit 610.
The camera 630 has a function of imaging reflected light from the inspection object 650 under the control of the control unit 610. The camera 630 may be either a (Charge Coupled Devices) CCD system, a (Complementary Metal Oxide Semiconductor) CMOS system, or another system camera.

  In addition, the camera 630 transmits the captured data to the control unit 610 as image data. Control unit 610 transmits the received image data to control unit 510 via communication unit 560. The control unit 510 stores the received image data in the storage unit 520 as a learning original image described later. Control unit 510 performs processing for displaying the received image data on display unit 530. An inspector (operator) can determine whether or not the inspection object 650 is a defective product by referring to an image (hereinafter also referred to as an inspection image) displayed on the display unit 530. Note that the inspector (operator) may directly check the inspection object 650 and determine whether or not the product is defective.

  Here, the terms used below are defined. First, a genuine defect that causes an inspector (operator) to determine that the inspection object 650 is defective is regarded as a true defect. On the other hand, although it is a defect extracted in the inspection process, the inspector (operator) does not reach the inspection object 650 as a defective product, and the defect that can be determined as the non-defective product is a pseudo defect. And

  Next, among the images obtained by imaging the inspection object 650, an image in which no true defect exists in the image regardless of the presence or absence of a pseudo defect is defined as a non-defective image, and the true defect exists in the image regardless of the presence or absence of a pseudo defect. At least one existing image is defined as a defect image. The non-defective image and the defect image can be learning images.

  Assume that the size of an image used in the following embodiments is, for example, horizontal 640 pixels and vertical 480 pixels. Note that the size of an image used in the following embodiments is not limited to 640 pixels wide and 480 pixels high, and may be any size (eg, 320 pixels wide and 240 pixels high).

  Further, the density values of the pixels of the image used in the following embodiments are values in the range of “0” to “255”. If the density value is “0”, the corresponding pixel is assumed to be black. If the density value is “255”, the corresponding pixel is white. Note that the density values of the pixels of the image used in the following embodiments are not limited to values in the range of “0” to “255”, and may be values in an arbitrary range.

  FIG. 3 is a functional block diagram of learning device 500 in the present embodiment. Referring to FIG. 3, an image data input unit 600 and a display unit 530 are connected to learning device 500.

  The learning device 500 includes a control unit 510. The control unit 510 operates as a preprocessing unit 511, a feature extraction unit 512, an identification unit 513, an erroneous specification determination unit 513A, a calculation unit 514, and an evaluation unit 515.

  The preprocessing unit 511 performs a process to be described later as a preprocessing unit. The feature extraction unit 512 performs processing to be described later as feature extraction means. The identification part 513 performs the process mentioned later as an identification means. The erroneous designation determination unit 513A performs the process described later as an erroneous designation determination unit. The calculation part 514 performs the process mentioned later as a calculation means. The evaluation part 515 performs the process mentioned later as an evaluation means.

  FIG. 4 is a diagram illustrating a learning image 700 as an example. Referring to FIG. 4, learning image 700 includes true defects 720, 722, 724 and pseudo defects 710, 712, 714. The area other than the true defects 720, 722, 724 and the pseudo defects 710, 712, 714 of the learning image 700 is the background area. That is, the learning image 700 includes a background area and a defect area.

  In the learning image, an inspector (operator) designates a defect whose size, shape, position, etc. are set as a true defect or a pseudo defect, and does not give a defect image or a true defect to which the defect is given. It is a good product image. Further, the learning image may be an image obtained by imaging the inspection object 650 regardless of the presence or absence of a true defect. The storage unit 520 stores a plurality of different types of learning images.

  Here, the learning image is an image subjected to image processing by an image processing algorithm.

  The learning apparatus 500 performs a plurality of types of learning images that are different from each other, and performs image processing algorithm generation processing that learns and generates the best image processing algorithm that is not detected and overdetected. The learning image is an image including a plurality of learning patterns or an image not including a learning pattern. The learning apparatus 500 generates the best image processing algorithm so that defects as learning patterns included in the learning image can be classified into two classes (a true defect class and a pseudo defect class).

  The image processing algorithm is, for example, an algorithm that sequentially uses a plurality of image processing modules. Each image processing module has several parameters.

  Here, the image processing module is, for example, a maximum value filter. The maximum value filter is a filter that performs processing for replacing the pixel value of interest with the maximum luminance value of the pixel of interest and the pixels in the vicinity thereof. In addition, as a parameter of the maximum value filter, for example, a parameter that defines a neighborhood region is included. When the value of the parameter is, for example, “3”, the filter size is defined as 3 pixels × 3 pixels. In this case, the target pixel and 8 pixels in the vicinity thereof are targeted for filtering. Similarly, when the parameter value is “5”, the filter size is defined as 5 pixels × 5 pixels. In this case, the pixel of interest and 24 pixels in the vicinity of the pixel of interest are subject to filter processing. The image processing algorithm is expressed by a combination of various image processing modules and a parameter value possessed by each image processing module.

  The image processing algorithm generation process repeats a process of changing a combination of image processing modules constituting the image processing algorithm and a parameter value of each image processing module based on an evaluation value calculated by a calculation method described later. This is a method for generating an image processing algorithm having a higher evaluation value.

  The one or more image processing modules constituting the image processing algorithm are selected from those previously generated and registered by an operator or a program. Specifically, the generated one or more image processing modules are stored in the storage unit 520.

FIG. 5 is a flowchart of the image processing algorithm generation process.
Referring to FIG. 5, in step S300, pre-processing unit 511 generates an initial image processing algorithm. Specifically, the preprocessing unit 511 generates an image processing algorithm by combining an image processing module group and parameters arbitrarily selected from the image processing module group stored in the storage unit 520 and its parameter group. Thereafter, the process proceeds to step S310.

  In step S <b> 310, the preprocessing unit 511 reads one learning image from the storage unit 520. Thereafter, the process proceeds to step S315.

  In step S315, image processing is performed. In the image processing, the preprocessing unit 511 performs image processing based on the generated image processing algorithm on the learning image read in step S310, and acquires a processed image. The image processing is processing for enhancing the target part (true defect) of the learning image and weakening the non-target part (pseudo defect). In the image processing, noise removal and the like are also performed. Thereby, the detection accuracy of true defects is increased. Thereafter, the process proceeds to step S340.

  In step S340, a feature amount calculation process is performed. In the feature amount calculation process, the feature extraction unit 512 extracts all defect candidate portions corresponding to the true defect or the pseudo defect included in the processed image, and calculates the value of each feature amount of all the defect candidate portions. Hereinafter, the data indicating the feature value is also referred to as feature data.

  Here, the feature amount is, for example, the contrast between the defect candidate portion and the vicinity thereof, the area of the defect candidate portion, shape information (perimeter length, width, number of holes, Euler number, moment amount, etc.), barycentric coordinates, color information, etc. It is. Thereafter, the process proceeds to step S350.

  In step S350, defect identification processing is performed. In the defect identification process, the identification unit 513 classifies each defect candidate part extracted in step S340 into a true defect and a pseudo defect. That is, the defect candidate part is classified into two classes.

  First, normal separation processing is performed to correctly classify defect candidate portions. In the normal separation process, the identification unit 513 determines all defect candidate parts based on the comparison between the position data of each defect candidate part and the position data of the true defect or the pseudo defect previously given to the learning image by the operator. Classify as true defect or pseudo defect. The classification method is, for example, a method in which a defect candidate part closest to each true defect is regarded as a true defect and other defect candidate parts are regarded as pseudo defects.

  Hereinafter, a class to which a defect candidate portion classified as a true defect belongs is also referred to as a true defect class. A class to which a defect candidate part classified as a pseudo defect belongs is also referred to as a pseudo defect class.

  Next, pattern recognition classification processing is performed. The pattern recognition classification process is performed based on the calculated feature value data of each defect candidate portion. In the following, a case where there is one type of feature amount will be described. In the pattern recognition classification process, the identification unit 513 stores feature value data threshold values (hereinafter also referred to as class classification threshold values), the feature value data values of the defect candidate portions to be classified, and the class classification. The magnitude relationship with the threshold value is determined, and classified into a true defect and a pseudo defect. For example, a defect candidate portion corresponding to feature amount data equal to or higher than the class classification threshold is classified as a true defect, and a defect candidate portion corresponding to feature amount data less than the threshold is classified as a pseudo defect. That is, the class classification threshold is a value for classifying the defect candidate portion into a pseudo defect and a true defect.

  If the defect candidate portion classified as a true defect by the normal separation process is classified as a true defect by the pattern recognition classification process, it means that the identification unit 513 has correctly detected the true defect. On the other hand, when the defect candidate part classified into the pseudo defect by the normal separation process is classified into the pseudo defect by the pattern recognition classification process, the identification unit 513 was correctly detected as the pseudo defect (it was not detected as the true defect). )It turns out that. In this case, the defect candidate portion to be classified as a true defect is correctly classified as a true defect by the identification unit 513. In addition, the defect candidate portion to be classified as a pseudo defect is correctly classified as a pseudo defect by the identification unit 513.

  On the other hand, if the defect candidate part classified as a true defect by the normal separation process is classified as a pseudo defect by the pattern recognition classification process, the identification part 513 cannot correctly detect the true defect (hereinafter, not detected). Also called).

  In addition, when a defect candidate part classified as a pseudo defect by the normal separation process is classified as a true defect by the pattern recognition classification process, the identification unit 513 incorrectly classifies the pseudo defect as a true defect (hereinafter, excessive defect). Also called detection).

  It is desirable that no detection or over detection be zero. Therefore, the learning identifying unit 513 may subtract a predetermined value from the evaluation value of the image processing algorithm according to the number of occurrences when undetected or overdetected. That is, the evaluation value improves as the number of undetected or overdetected items decreases.

  Various methods have been proposed in the past as a method for the classifier (identification unit 513) to perform pattern recognition classification. A method of discriminating feature amount data based on a threshold value as described above is called a method using a discrimination function. There are other methods such as a k-NN method (k-Nearest-Neighbours Classification Rule) and a method using a neural network. In the learning apparatus of the present invention, any classifier method may be used. Thereafter, the process proceeds to step S352.

  In step S352, the erroneous specification determination unit 513A determines whether or not a counter K, which will be described later, is greater than or equal to a predetermined value (for example, 50). If YES at step S352, control proceeds to step S354. On the other hand, if NO at step S352, the process proceeds to step S360.

  In step S352, it may be determined whether a predetermined time has elapsed since the process of step S300 was started, instead of the value of counter K.

  In step S354, the erroneous specification determination unit 513A determines whether there is an erroneous specification. Here, the erroneous designation means that there has been no detection or over detection in step S350. That is, when an operator generates a learning image, there is a high possibility that a place where a defect should be designated as a true defect (or a pseudo defect) is mistakenly designated as a false defect (or a true defect). If YES in step S354, the process proceeds to step S356. On the other hand, if NO at step S354, the process proceeds to step S360.

  In step S356, the erroneous designation determination unit 513A excludes feature amount data corresponding to a defect candidate portion that has not been detected or overdetected from a target to be used in a calculation that will be described later. In order to notify the operator of the exclusion, the display unit 530 displays that fact. Thereafter, the process proceeds to step S360.

  In this case, based on the display on the display unit 530, the operator may generate a learning image in which a defect that has been incorrectly specified is correctly specified. Then, the processing may be performed again from step S300 of the image processing algorithm generation processing.

  In step S360, a distribution state calculation process is performed. In the distribution state calculation process, the calculation unit 514 includes the feature quantity distribution of the defect candidate part classified into the pseudo defect class and the feature quantity distribution of the defect candidate part classified into the true defect class in the pattern recognition classification process described above. Based on the above, the intra-class dispersion degree and the inter-class dispersion degree are calculated.

  The intra-class dispersion degree is an index representing the degree of dispersion from the intra-class average with respect to the feature amount distribution of the defect candidate portion classified into the pseudo defect class or the true defect class. In the following, the intra-class dispersion degree corresponding to the pseudo defect class is also referred to as the pseudo-defect class dispersion degree SW (1). Further, the intra-class dispersion degree corresponding to the true defect class is also referred to as a true defect class dispersion degree SW (2).

  The inter-class dispersion degree is an index representing the average degree of dispersion within two classes. In the following, an index representing the average degree of dispersion within each of the pseudo defect class and the true defect class is also referred to as an inter-class dispersion degree SB. Details of the calculation method of the intra-class dispersion degree and the inter-class dispersion degree, which is a feature of the present invention, will be described later. Thereafter, the process proceeds to step S370.

  In step S370, the evaluation unit 515 calculates the evaluation value E of the image processing algorithm used in step S315 based on the intra-class dispersion degree and the inter-class dispersion degree SB calculated in step S360 by the following equation (1). Calculate based on

According to Equation (1), it can be seen that the evaluation value E is calculated by dividing the inter-class dispersion degree SB by the total value of all the intra-class dispersion degrees.

  According to Equation (1), the evaluation value E increases as the scatter degree SB within the class decreases. Further, the evaluation value E increases as the degree of dispersion between classes increases.

  The evaluation value E may be calculated only from the degree of dispersion within the class. Further, the evaluation value E may be calculated only from the degree of dispersion between classes. However, the evaluation value E is preferably calculated from both the intra-class dispersion degree and the inter-class dispersion degree.

  Formula (1) is a formula for calculating an evaluation value when a class is classified into two. In the case of classifying into three or more classes, the evaluation value E is calculated by dividing the total dispersion degree between all classes by the total dispersion degree within all classes. Thereafter, the process proceeds to step S375.

  In step S375, the evaluation unit 515 determines whether an end condition is satisfied. Here, the end condition in the present embodiment is that the evaluation value calculated in step S370 is equal to or greater than a predetermined value set in advance, or the number of times the image processing algorithm parameter adjustment processing described later is repeated. The condition is a predetermined number of times (for example, 100 times). The termination condition may be other conditions.

  If YES in step S375, the image processing algorithm generation process ends. The image processing algorithm generation process ends when, for example, the evaluation value calculated in step S370 is greater than or equal to a predetermined value. Or it is a case where the frequency | count of repeating the parameter adjustment process of the image processing algorithm mentioned later becomes predetermined times.

On the other hand, if NO at step S375, the process proceeds to step S380.
In step S380, the evaluation unit 515 changes (adjusts) the combination of the image processing modules and the parameter values so that the evaluation value becomes large based on the evaluation value calculated for the generated image processing algorithm. The changed image processing algorithm is used in step S315.

  As a technique for optimizing a combination of image processing modules and parameter values based on a certain evaluation value in step S380, for example, GA (genetic algorithm), SA (simulated annealing), nonlinear programming, or neural network is used. Is. In either case, for combinations and parameter values with high evaluation values, those combinations and values are slightly changed to search for ones with higher evaluation values.

  If the number of combinations and the range of parameter values are sufficiently small, the number and parameter values of all combinations of image processing modules that are components of the image processing algorithm are within the range that does not satisfy the termination condition of step S375. It is possible to perform evaluation under varying conditions. When the adjustment of the image processing algorithm is completed as described above, the process proceeds to step S382.

  In step S382, the evaluation unit 515 increments the counter K by 1. Here, the counter K is a counter for counting the number of times the process of step S380 has been performed. Note that the initial value of the counter K is set to zero. Thereafter, the process of step S315 is performed again.

  That is, the processes in steps S310, S315, S340, S350, S360, S370, and S380 are processes for changing the image processing algorithm.

  By repeatedly performing the processes of steps S315, S340, S350, S360, S370, and S380 until the end condition of step S375 is satisfied, generation of a final image processing algorithm having an evaluation value equal to or greater than a predetermined value can be realized.

  Next, a method for calculating the intra-class dispersion degree and the inter-class dispersion degree will be described in detail. As described above, the calculation unit 514 is based on the feature amount distribution of the defect candidate portion classified into the pseudo defect class and the feature amount distribution of the defect candidate portion classified into the true defect class in the pattern recognition classification process described above. Then, the intra-class dispersion degree and the inter-class dispersion degree are calculated.

  FIG. 6 is a diagram illustrating a first example of the feature amount distribution (histogram). FIG. 6 is a plot diagram generated based on the 110 feature data acquired by the process of step S340. That is, the learning image read in step S310 has 110 defect candidate portions. In the example of FIG. 6, the feature amount data is only one type. The feature amount data is a contrast value that takes a value in the range of 0 to 200. The contrast in this case is a value obtained by quantifying the contrast between the defect and the background of the surrounding area. The feature amount data is not limited to the contrast, and may be a defect area or the like.

  It is assumed that 100 pieces of feature data out of 110 pieces of feature data are data classified into pseudo defect classes. Further, it is assumed that 10 pieces of feature amount data out of 110 pieces of feature amount data are data classified into true defect classes.

  FIG. 7 is a diagram illustrating the feature amount data used in the first example of the feature amount distribution. FIG. 7A is a diagram showing a data table T100 composed of 100 pieces of feature amount data classified into the pseudo defect class. FIG. 7B is a diagram showing a data table T110 composed of ten pieces of feature data classified into true defect classes.

  Referring to FIG. 6 again, the horizontal axis indicates the value of the feature amount data, and the vertical axis indicates the frequency. Here, the frequency indicates the number of defect candidate parts whose corresponding feature value data belongs to a value in a predetermined range. In FIG. 6, for example, when the frequency is “21”, it indicates that the number of defect candidate parts belonging to a value range in which the value of the corresponding feature amount data is 45 or more and less than 55 is 21.

  In FIG. 6, the class region C1A includes a curve based on data classified into a pseudo defect class. The class area C1B includes a curve based on data classified into the true defect class. In FIG. 6, there is no exceptional value far away from the curve of the largest mountain in each class area. In this case, there is no problem even if the arithmetic average is calculated as the intra-class average and the variance is calculated as the intra-class dispersion degree.

  FIG. 8 is a diagram illustrating a second example of the feature amount distribution (histogram). Since the horizontal and vertical axes are the same as those in FIG. 6, detailed description will not be repeated. FIG. 8 is a plot diagram generated based on 111 pieces of feature amount data acquired by the process of step S340. That is, the learning image read in step S310 has 111 defect candidate portions. The feature amount data is a contrast value that takes a value in the range of 0 to 200 as in FIG. In the example of FIG. 8, as in FIG. 6, there is only one type of feature amount data.

  It is assumed that 100 pieces of feature data out of 111 pieces of feature data are data classified into pseudo defect classes. Further, it is assumed that 11 pieces of feature amount data among 111 pieces of feature amount data are data classified into true defect classes.

  FIG. 9 is a diagram illustrating the feature amount data used in the second example of the feature amount distribution. FIG. 9A is a diagram showing a data table T100 composed of 100 pieces of feature amount data classified into the pseudo defect class. FIG. 9B is a diagram showing a data table T110A composed of eleven feature quantity data classified into the true defect class.

Referring to FIG. 8 again, class region C2A includes a curve based on data classified into a pseudo defect class. The class area C2B includes a curve based on data classified into the true defect class. Compared with FIG. 6, FIG. 8 includes one piece of data with a feature value data value of 150 classified into a true defect class. The data is far from the largest peak curve in the class area C2B, and becomes an exceptional value. Excluding the exceptional value, the feature amount distributions of the first example (FIG. 6) and the second example (FIG. 8) are exactly the same.
(Examination example of calculation method)
FIG. 10 is a diagram illustrating a calculation result as an examination example by the calculation unit 514 and the evaluation unit 515 based on the first and second examples of the feature amount distribution.

  Referring to FIG. 10, m (1) indicates an intraclass average based on data in class area C1A and class area C2A. SW (1) indicates the degree of dispersion within the class based on the data in the class area C1A and the class area C2A.

  m (2) represents an intraclass average based on data in the class area C1B and the class area C2B. SW (2) indicates the degree of dispersion in the class based on the data in the class area C1B and the class area C2B.

  SB indicates the degree of inter-class dispersion, which is an index representing the average degree of dispersion within two classes. E is an evaluation value of the image processing algorithm.

  The in-class average m (1) in the first example is expressed by the following equation (2) for calculating the arithmetic average based on 100 feature data (see FIG. 7A) in the class region C1A. 40.3.

N (i) in Equation (2) is the number of feature quantity data of class i. In this case, i is “1”. Therefore, N (1) is 100. Sij is the value of j-th data (feature data) of class i.

  The in-class average m (1) in the second example is calculated by Expression (2) based on 100 feature data in the class area C2A. The 100 feature quantity data in the class area C2A (see FIG. 9A) is the same as the 100 feature quantity data in the class area C1A. 1) is the same value (40.3) as the in-class average m (1) in the first example.

The fluctuation rate is calculated by the following equation (3).
[(Value of second example) / (value of first example) -1] × 100 (3)
Therefore, the variation rate of the in-class average m (1) in the first and second examples is “0” from the equation (3).

  The intra-class dispersion degree SW (1) in the first example is calculated based on the following equation (4) and the above-described equation (2) based on 100 feature data in the class region C1A. Calculated as 37.0.

In this case, i in the formula (4) is “1”. Note that N (i) -1, which is the denominator of the formula (4), may be N (i).

  The intra-class dispersion degree SW (1) in the second example is calculated by the equations (4) and (2) based on 100 feature data in the class region C2A. Note that the 100 feature value data in the class region C2A (see FIG. 9A) is the same as the 100 feature value data in the class region C1A, and therefore, the intra-class dispersion degree SW in the second example. (1) is the same value (37.0) as the intra-class dispersion degree SW (1) in the first example.

  Further, the variation rate of the intra-class dispersion degree SW (1) in the first and second examples is “0” from the equation (3).

  The in-class average m (2) in the first example is expressed by the following equation (2) with i = 2, N (i) based on 10 feature quantity data (see FIG. 7B) in the class area C1B. ) = 10, 97.7 is calculated.

  The in-class average m (2) in the second example is expressed by the following equation (2) based on 11 pieces of feature amount data (see FIG. 9B) in the class region C2B. ) = 11, it is calculated as 102.4.

Further, the variation rate of m (2) in the first and second examples is “5” from the equation (3).
The intra-class dispersion degree SW (2) in the first example is calculated based on the ten feature quantity data in the class region C1B, with i = 2, N (i) = 10, By substituting m (2) in the example of 1, it is calculated as 20.9.

  The intra-class dispersion degree SW (2) in the second example is calculated based on 11 feature value data in the class region C2B, i = 2, N (i) = 11, By substituting m (2) in the example of 2, it is calculated as 267.8.

  Further, the variation rate of the intra-class dispersion degree SW (2) in the first and second examples is “1181” from the equation (3).

  The inter-class dispersion degree SB in the first example is calculated as 821.7 by the following expressions (5) and (6).

  In the equation (6), μ is an average value of the in-class average m (1) and the in-class average m (2).

  The inter-class dispersion degree SB in the second example is calculated as 963.8 by the equations (5) and (6).

  Further, the variation rate of the inter-class dispersion degree SB in the first and second examples is “17” from the equation (3).

  The evaluation value E of the image processing algorithm in the first example is calculated as 14.2 from the above-described equation (1). It can be seen from FIG. 6 that the 110 feature quantity data in the first example can be separated into the pseudo defect class and the true defect class in the vicinity of 80 of the feature quantity data. Therefore, it can be easily understood that an excellent image processing algorithm without undetected or overdetected has been generated.

  Note that the evaluation value E of the image processing algorithm in the second example is calculated to be 3.16 from the above equation (1). As described above, the evaluation value E of the image processing algorithm includes an exceptional value in the feature amount data in the second example.

  Further, the variation rate of the evaluation value E in the first and second examples is “−78” from the equation (3).

  Therefore, it can be seen that the fact that there is only one exceptional value greatly affects the intra-class average, the intra-class dispersion degree SW, the inter-class dispersion degree SB, and the evaluation value E.

For example, the intraclass average m (2) of true defects in the first and second examples varies by 5% from 97.7 to 102.4. Also, the true defect intra-class dispersion degree SW in the first and second examples greatly fluctuates by 1181% from 20.9 to 267.8. Further, the true defect inter-class dispersion degree SB in the first and second examples varies by 17% from 821.7 to 963.8. In addition, the evaluation value E of the image processing algorithm in the first and second examples is greatly changed by -78% from 14.2 to 3.16, and the evaluation is lowered.
Thus, in both the first and second examples, the average of the intraclass average is used in spite of the performance of the image processing algorithms used to classify them into the pseudo defect class and the true defect class. A problem arises in the calculation method of the study example in which the calculation is performed in step (b) and the dispersion degree SW in the class is calculated by dispersion.

(Calculation method 1 in the first embodiment)
Next, a calculation method for reducing the influence of exceptional values will be described. As an example, if M-estimator, which is a technique for correctly estimating the average value using weighting, is used, the in-class average, the in-class dispersion degree, and the evaluation value can be accurately calculated even when an exceptional value is included. The intraclass average using M-estimator is calculated by the following formula (7).

In Expression (7), K is an arbitrary variable that is a candidate for intra-class average (hereinafter also referred to as an intra-class average candidate variable). m (i) is an in-class average of class i, and argmin {x} is an expression obtained by changing K so that x becomes a minimum value.

  In Expression (7), N (i) is the number of defect candidate parts (feature data) belonging to class i, and f (i, j) is the jth feature data of the defect candidate parts belonging to class i. Indicates.

  The function ρ is called a weighting error function, and various proposals have been made so far. For example, the Geman-McClure function is shown in the following formula (8).

  Here, σ is a predetermined constant, and is a parameter that determines how much weighting is performed in accordance with the distance from the average (in this case, the intraclass average m (i)). In this way, the intraclass average m (i) can be calculated by using the M-estimator.

  In the present invention, it is proposed to use the M-estimator to calculate the intra-class dispersion degree SW by the following equation (9).

In equation (9), SW (i) is the degree of dispersion within class i, and m (i) is the class average of class i calculated in equation (7).

  FIG. 11 is a diagram illustrating a calculation result by the calculation method 1 of the present invention by the calculation unit 514 and the evaluation unit 515 based on the first and second examples of the feature amount distribution. Note that the feature data used in the first and second examples is the same as the data used in the first and second examples described above. In the following calculation, σ in equation (8) is 6.

  Referring to FIG. 11, in-class average m (1) in the first example is calculated by equation (7) based on 100 feature quantity data (see FIG. 7A) in class region C1A. Value. The in-class average m (1) in the second example is a value calculated by Expression (7) based on 100 feature data in the class area C2A.

  The intra-class dispersion degree SW (1) in the first example is a value calculated by the proposed formula (9) based on 100 feature quantity data in the class area C1A. The intra-class dispersion degree SW (1) in the second example is a value calculated by the proposed equation (9) based on 100 feature data in the class region C2A.

  The in-class average m (2) in the first example is a value calculated by the equation (7) based on the ten feature amount data (see FIG. 7B) in the class region C1B. The in-class average m (2) in the second example is a value calculated by Expression (7) based on 11 pieces of feature data (see FIG. 9B) in the class area C2B. The intra-class dispersion degree SW (2) in the first example is a value calculated by the proposed equation (9) based on the ten feature quantity data in the class area C1B. The intra-class dispersion degree SW (2) in the second example is a value calculated by the proposed formula (9) based on 11 pieces of feature amount data in the class region C2B.

  The inter-class dispersion degree SB in the first and second examples is a value calculated by the equations (5) and (6). The evaluation value E of the image processing algorithm in the first and second examples is a value calculated by Expression (1).

  In-class average m (2) of true defects in the first and second examples was affected by the exceptional value in the study example, but according to the calculation method 1 described above, without being affected by the exceptional value, It has not changed at all. Accordingly, the true defect inter-class dispersion degree SB in the first and second examples is not affected by the exceptional value and does not change at all.

  In the first and second examples, the true defect intra-class dispersion degree SW (2) was greatly affected by the exceptional value in the study example. The fluctuation is only 28%, from 25 to 0.32.

  Therefore, the evaluation value E of the image processing algorithm in the first and second examples was also greatly affected by the exceptional value in the study example, but according to the calculation method 1, from 1329.9 to 1192.2, The fluctuation is only -10%.

  Therefore, by using the calculation method based on the proposed equation (9) using the M-estimator, the influence of the exceptional value can be reduced, and the image processing algorithm can be evaluated correctly.

  Therefore, also in the above-described step S340, the combination of image processing modules and the change (adjustment) of parameter values can be more accurately performed based on the evaluation value of the image processing algorithm that has been correctly evaluated. That is, there is an effect that learning can be performed more accurately.

  Therefore, it is possible to generate an image processing algorithm for increasing the defect identification rate in a short time. There is an effect.

  The calculation method when the feature amount data is one type has been described above, but the present invention using the M-estimator can be implemented even when there are a plurality of types of feature amount data. In this case, a plurality of types of feature amount data of each of the plurality of defect candidate portions may be represented by feature vectors and extended from scalar calculation to vector calculation. The intraclass average m (i) of class i can be calculated by the following equation (10).

Here, || x, y || represents the distance between the vector x and the vector y. For example, the Euclidean distance may be used as the distance. K is an intraclass average candidate variable. Further, the intra-class dispersion degree SW of class i can be calculated by the following formula (11) proposed in the present invention.

Here, m (i) in equation (11) is the intraclass average m (i) of class i calculated in equation (10).

(Calculation method 2 in the first embodiment)
Next, another calculation method for reducing the influence of the exceptional value will be described. As an example, if LMedS (Least Median of Squares), which is a method using a median value as an average value, is used, the average within the class, the degree of dispersion within the class, and the evaluation value can be accurately calculated even when exceptional values are included. . The intraclass average using LMedS is calculated by the following equation (12).

In Expression (12), Med {x ² } is a square distribution of deviation x of a j ((j = 1,..., N (i)))-th defect candidate portion (feature data) belonging to class i. Is the median. K is an intraclass average candidate variable. x is a distance (difference) between the feature amount data f (i, j) of each of the plurality of defect candidate portions belonging to the class i and the in-class average candidate variable K. As described above, in the method of using the median when estimating the average value, the weight by the exceptional value is reduced, and the dispersion degree within the class is calculated using the weight.

  Furthermore, in the present invention, it is proposed to calculate the intra-class dispersion degree SW using the following formula (13) using LMedS.

In equation (13), m (i) is the intraclass average of class i calculated in equation (12).

  FIG. 12 is a diagram illustrating a calculation result by the calculation method 2 of the present invention by the calculation unit 514 and the evaluation unit 515 based on the first and second examples of the feature amount distribution. Note that the feature data used in the first and second examples is the same as the data used in the first and second examples described above.

  Referring to FIG. 12, the in-class average m (1) in the first example is calculated by Expression (12) based on 100 feature data (see FIG. 7A) in the class area C1A. Value. The in-class average m (1) in the second example is a value calculated by Expression (12) based on 100 feature quantity data in the class area C2A.

  The intra-class dispersion degree SW (1) in the first example is a value calculated by the proposed formula (13) based on 100 feature data in the class region C1A. The intra-class dispersion degree SW (1) in the second example is a value calculated by the proposed formula (13) based on 100 feature quantity data in the class area C2A. The in-class average m (2) in the first example is a value calculated by the equation (12) based on 10 feature quantity data (see FIG. 7B) in the class area C1B. The in-class average m (2) in the second example is a value calculated by Expression (12) based on 11 pieces of feature data (see FIG. 9B) in the class area C2B. The intra-class dispersion degree SW (2) in the first example is a value calculated by the proposed formula (13) based on the ten feature quantity data in the class area C1B. The intra-class dispersion degree SW (2) in the second example is a value calculated by the proposed equation (13) based on the 11 feature data in the class area C2B.

  The inter-class dispersion degree SB in the first and second examples is a value calculated by the equations (5) and (6). The evaluation value E of the image processing algorithm in the first and second examples is a value calculated by Expression (1).

  In-class average m (2) of true defects in the first and second examples was affected by the exceptional value in the examination example, but according to the above calculation, it was completely affected by the exception value. Not done. Accordingly, the true defect inter-class dispersion degree SB in the first and second examples is not affected by the exceptional value and does not change at all.

  In the first and second examples, the true defect intra-class dispersion degree SW is greatly influenced by the exceptional value in the study example, but according to the calculation according to the proposed equation (13), 9.1 to 9 It is only 1% of fluctuation.

  Therefore, the evaluation value E of the image processing algorithm in the first and second examples is also greatly affected by the exceptional value in the study example, but according to the calculation method 2, the evaluation value E is not affected by the exceptional value. It has not changed at all.

  Therefore, by using the calculation method according to the proposed equation (13) using LMedS, the influence of the exceptional value can be reduced, and the image processing algorithm can be evaluated correctly.

  Therefore, also in the above-described step S340, the combination of image processing modules and the change (adjustment) of parameter values can be more accurately performed based on the evaluation value of the image processing algorithm that has been correctly evaluated. That is, there is an effect that learning can be performed more accurately.

  Therefore, it is possible to generate an image processing algorithm for increasing the defect identification rate in a short time. There is an effect.

  Although the calculation method in the case where there is one type of feature amount data has been described above, the present invention using LMedS can be implemented even when there are a plurality of types of feature amount data. In this case, a plurality of types of feature amount data of each of the plurality of defect candidate portions may be represented by feature vectors and extended from scalar calculation to vector calculation. The intraclass average m (i) of class i can be calculated by the following equation (14).

  Furthermore, in the present invention, it is proposed to calculate the intra-class dispersion degree SW using the following formula (15) using LMedS.

In equation (15), m (i) is the intraclass average of class i calculated in equation (14).

  In the first embodiment, the evaluation value is fed back to the preprocessing unit 511 to learn the image processing algorithm. In the present invention, the evaluation value may be fed back to the feature extraction unit 512 or the identification unit 513 to be learned. Learning the feature extraction unit 512 is, for example, selecting an appropriate feature amount. Learning the identification unit 513 means, for example, learning a class classification threshold for separating a defect candidate part into a pseudo defect and a true defect.

  In the second example shown in FIG. 8, there is a case where there is feature data whose exception value is far away from the curve of the largest mountain in each class region. However, exceptional values may exist in other class distributions besides the class to which they belong.

  In other words, this is a case where the operator has mistakenly designated a true defect, despite the defect candidate portion originally belonging to the pseudo defect class. Or, it is a case where the operator overlooks the designation as a true defect regardless of the defect candidate portion originally belonging to the true defect class.

  In this case, it is not possible to generate a learning algorithm that completely separates the pseudo defect and the true defect (that is, sets both overdetection and undetection to zero). Even in the case where such an exceptional value exists, the present invention can calculate an evaluation value so as not to be affected by the exceptional value, and has an effect that a learning algorithm can be learned.

<Second Embodiment>
Next, a process for performing pattern recognition using the image processing algorithm generated in the first embodiment will be described.

  FIG. 13 is a block diagram showing an internal configuration of pattern recognition apparatus 500A in the present embodiment.

  Referring to FIG. 13, pattern recognition apparatus 500 </ b> A has data access to recording medium 555 </ b> A instead of recording medium 555 as compared with learning apparatus 500 of FIG. The difference is that the program 180A is recorded, the program 180A is recorded on the recording medium 555A, and the image processing algorithm 184 is further recorded in the storage unit 520. Since other configurations and functions are similar to those of learning apparatus 500, detailed description will not be repeated.

  The image processing algorithm 184 is an image processing algorithm generated in the first embodiment. Here, the image processing algorithm is composed of one or more image processing filters. The pattern recognition apparatus 500A receives a processing target image for identifying a pattern in the image. Information about one or more defects existing in the processing target image is unknown.

  The pattern recognition apparatus 500A can detect each defect individually and estimate the class to which the defect belongs. Of course, when image data having no defect is input, the pattern recognition apparatus 500A outputs an estimation result that there is no defect to be detected. Pattern recognition processing for recognizing the pattern of the input image is performed by the pattern recognition apparatus 500A having the above configuration.

Next, the pattern recognition process will be described.
FIG. 14 is a flowchart of pattern recognition processing. Referring to FIG. 14, in step S <b> 410, control unit 510 reads image processing algorithm 184 stored in storage unit 520 from storage unit 520. Thereafter, the process proceeds to step S420.

  In step S420, a processing target image is input by imaging the inspection object 650 using the image data input unit 600. Further, when performing offline image inspection, an image is input by reading the processing target image stored in the storage unit 520. Thereafter, the process proceeds to step S430.

  In step S430, image processing is performed. In the image processing, the preprocessing unit 511 generates a processed image obtained by performing the processing of the image processing algorithm 184 on the processing target image acquired in step S420. Thereafter, the process proceeds to step S440.

  In step S440, a feature amount calculation process is performed. In the feature amount calculation processing, the feature extraction unit 512 extracts all defect candidate portions corresponding to the true defect or the pseudo defect included in the processed image, and calculates the feature amount data value corresponding to each of all the defect candidate portions. calculate. Thereafter, the process proceeds to step S450.

  In step S450, defect identification processing is performed. In the defect identification process, the identification unit 513 classifies each defect candidate part extracted in step S340 into a true defect and a pseudo defect. That is, the defect candidate part is classified into two classes. Since the classification method is the same as the pattern recognition classification process described above, detailed description will not be repeated. Therefore, by classifying each defect candidate part into a true defect and a pseudo defect, the defect candidate part as a pattern in the processing target image is identified. Thereafter, the process proceeds to step S460.

  In step S460, the classification result of each defect candidate part classified in step S450 is displayed on display unit 530. Thus, the pattern recognition process ends.

  As described above, in this embodiment, the image processing algorithm generated in the first embodiment for increasing the defect identification rate is used during the pattern recognition processing.

  Therefore, there is an effect that a defect candidate portion as a pattern in the processing target image can be identified more accurately.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

It is a block diagram which shows the internal structure of the learning apparatus in this Embodiment. It is a block diagram which shows the internal structure of an image data input part. It is a functional block diagram of the learning apparatus in this Embodiment. It is a figure which shows the learning image as an example. It is a flowchart of an image processing algorithm generation process. It is a figure which shows the 1st example of feature-value distribution (histogram). It is a figure which shows the feature-value data used for the 1st example of feature-value distribution. It is a figure which shows the 2nd example of feature-value distribution (histogram). It is a figure which shows the feature-value data used for the 2nd example of feature-value distribution. It is a figure which shows the calculation result as a study example by the calculation part and evaluation part based on the 1st and 2nd example of feature-value distribution. It is a figure which shows the calculation result by the calculation method 1 of this invention by the calculation part and evaluation part based on the 1st and 2nd example of feature-value distribution. It is a figure which shows the calculation result by the calculation method 2 of this invention by the calculation part and evaluation part based on the 1st and 2nd example of feature-value distribution. It is a block diagram which shows the internal structure of the pattern recognition apparatus in this Embodiment. It is a flowchart of a pattern recognition process.

Explanation of symbols

  180, 180A program, 182 learning image database, 184 image processing algorithm, 500 learning device, 500A pattern recognition device, 510 control unit, 520 storage unit, 530 display unit, 532 VDP, 555, 555A recording medium, 600 image data input Department.

Claims (26)

Pre-processing means for performing pre-processing on a plurality of learning patterns in which a class belonging in advance is designated;
Feature extraction means for extracting a feature amount from each of the plurality of learning patterns subjected to the preprocessing;
Identification means for classifying each of the plurality of learning patterns into one of a plurality of classes based on the value of the corresponding feature amount;
For each of the plurality of types of classes, the degree of scatter within the class indicating the distribution state in the corresponding class for the feature value corresponding to each of the plurality of learning patterns classified into each of the plurality of types of classes. Calculating means for calculating a degree of inter-class dispersion indicating a distribution state between the plurality of types of classes;
An evaluation means for calculating an evaluation value that satisfies one or both of the following: an evaluation value that increases as the degree of dispersion within the class decreases; and an evaluation value that increases as the degree of dispersion between classes increases. With
The evaluation unit learns at least one of the preprocessing unit, the feature extraction unit, or the identification unit by feeding back the calculated evaluation value so that the evaluation value becomes high,
The learning unit is configured to change the weighting when calculating the degree of dispersion in the class as the value of the feature quantity that is far from the range in which the feature quantity values for each of the plurality of types of classes are most distributed. The calculating means includes
Each of the plurality of types of learning patterns corresponding to each of the plurality of types of classes has a smaller weight as the value of the feature amount farther away from the range in which the value of the feature amount for each of the plurality of types of classes is the most distributed. An average in class based on the values of a plurality of corresponding feature amounts is calculated for each of the plurality of types of classes,
The degree of dispersion within the class representing the degree of dispersion from the average within the class is calculated for each of the plurality of types of classes,
The learning device according to claim 1, wherein the inter-class dispersion degree is calculated based on an intraclass average corresponding to each of the plurality of types of classes. The calculating means includes
A plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of the plurality of types of classes and an intra-class average candidate corresponding to the corresponding class are determined for each of the plurality of types of classes. To
For each of the plurality of distances, the weighted distance is weighted for each of the plurality of types of classes.
A plurality of weighted distances calculated are averaged based on the number of learning patterns corresponding to each of the plurality of types of classes, thereby calculating within-class dispersion degree candidates for each of the plurality of types of classes,
The learning according to claim 2, wherein the intra-class average candidate that minimizes the intra-class dispersion degree candidate is calculated as an intra-class average, and the intra-class dispersion degree candidate that is minimized is calculated as an intra-class dispersion degree candidate. apparatus.   The learning apparatus according to claim 3, wherein the averaging for calculating the degree of dispersion within the class by the calculating means is an arithmetic average. The calculating means includes
A plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of the plurality of types of classes and the average within the class corresponding to the corresponding class are determined for each of the plurality of types of classes. To
For each of the plurality of distances, a power distance that is a power of a predetermined power value is calculated for each of the plurality of types of classes,
The learning apparatus according to claim 2, wherein a median value based on the calculated distribution of the plurality of power distances is calculated for each of the plurality of types of classes as the degree of dispersion within the class.   The learning apparatus according to claim 5, wherein the power value is two.   The learning apparatus according to claim 3, wherein the distance is a Euclidean distance.   The evaluation means calculates, as the evaluation value, a value obtained by dividing the inter-class dispersion degree by a total value of a plurality of intra-class dispersion degrees corresponding to each of the plurality of types of classes. The learning device according to any one of 7. The identification means further comprises an erroneous specification determination means for determining that the learning pattern is classified into a class other than a previously assigned affiliation class as an erroneous specification,
The said incorrect designation | designated determination means excludes the learning pattern used as the object of the said erroneous designation from the learning pattern used as the object which calculates the said evaluation value which the said evaluation means calculates. The learning device described in 1. The learning pattern is an image,
The preprocessing is processing for performing image processing based on a predetermined image processing algorithm on the learning pattern,
The learning device according to claim 1, wherein the evaluation unit changes the image processing algorithm based on the evaluation value.   The learning device according to claim 10, wherein the evaluation unit changes the image processing algorithm based on a genetic algorithm. A step of pre-processing a plurality of learning patterns in which the belonging class is designated in advance;
Extracting a feature amount from each of the plurality of learning patterns subjected to the preprocessing;
Classifying each of the plurality of learning patterns into one of a plurality of classes based on the corresponding feature value;
For each of the plurality of types of classes, the degree of scatter within the class indicating the distribution state in the corresponding class for the feature value corresponding to each of the plurality of learning patterns classified into each of the plurality of types of classes. Calculating, and further calculating a degree of inter-class dispersion indicating a distribution state between the plurality of types of classes;
A step of calculating an evaluation value that satisfies any one or both of that the evaluation value increases as the degree of dispersion within the class decreases and that the evaluation value increases as the degree of dispersion between classes increases. Including
The step of calculating the evaluation value learns at least one of the step of performing the preprocessing, the step of extracting or the step of classifying by feeding back the calculated evaluation value so that the evaluation value becomes high Including the step of
The step of calculating includes a step of changing a weight when calculating the degree of dispersion within the class as the value of the characteristic amount that is far from the range in which the value of the characteristic amount for each of the plurality of types of classes is the most distributed. , Learning method. The step of calculating the inter-class dispersion degree includes
Each of the plurality of types of learning patterns corresponding to each of the plurality of types of classes has a smaller weight as the value of the feature amount farther away from the range in which the value of the feature amount for each of the plurality of types of classes is the most distributed. Calculating an average within a class based on a plurality of corresponding feature value values for each of the plurality of types of classes;
Calculating the degree of dispersion within the class representing the degree of dispersion from the average within the class for each of the plurality of types of classes;
The learning method according to claim 12, further comprising: calculating the degree of dispersion between classes based on an intraclass average corresponding to each of the plurality of types of classes. The step of calculating the inter-class dispersion degree includes
A plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of the plurality of types of classes and an intra-class average candidate corresponding to the corresponding class are determined for each of the plurality of types of classes. A step of calculating
For each of the plurality of distances, calculating a weighted distance for each of the plurality of types of weights, in which a weight to be calculated is smaller as each of the plurality of distances is larger.
A plurality of weighted distances calculated are averaged based on the number of learning patterns corresponding to each of the plurality of types of classes, thereby calculating within-class dispersion degree candidates for each of the plurality of types of classes,
And calculating the intra-class average candidate that minimizes the intra-class dispersion degree candidate as an intra-class average, and calculating the intra-class dispersion degree candidate that is minimized as the intra-class dispersion degree candidate. The learning method described in.   The learning method according to claim 14, wherein the step of calculating the inter-class dispersion degree is an arithmetic average in which the averaging for calculating the intra-class dispersion degree is an arithmetic average. The step of calculating the inter-class dispersion degree includes
A plurality of distances between a plurality of feature value values corresponding to each of a plurality of learning patterns corresponding to each of the plurality of types of classes and the average within the class corresponding to the corresponding class are determined for each of the plurality of types of classes. A step of calculating
For each of the plurality of distances, calculating a power distance that is a power of a predetermined power value for each of the plurality of types of classes;
The learning method according to claim 13, further comprising: calculating, for each of the plurality of types of classes, a median value based on the calculated distribution of the power distances as the degree of dispersion within the class.   The learning method according to claim 16, wherein the power value is two.   The learning method according to claim 14, wherein the distance is a Euclidean distance.   The step of calculating the evaluation value includes a step of calculating, as the evaluation value, a value obtained by dividing the inter-class dispersion degree by a total value of a plurality of intra-class dispersion degrees corresponding to each of the plurality of classes. The learning method according to any one of claims 12 to 18. The step of classifying further includes the step of determining that the learning pattern is classified into a class other than a pre-designated affiliation class as an erroneous designation,
The step of determining to be erroneously specified includes the step of excluding the learning pattern that is the target of erroneous specification from the learning pattern that is the target of calculating the evaluation value calculated by the step of calculating the evaluation value. The learning method according to claim 12. The learning pattern is an image,
The preprocessing is processing for performing image processing based on a predetermined image processing algorithm on the learning pattern,
21. The learning method according to claim 12, wherein the step of calculating the evaluation value includes a step of changing the image processing algorithm based on the evaluation value.   The learning method according to claim 21, wherein the step of calculating the evaluation value includes a step of changing the image processing algorithm based on a genetic algorithm. A learning program for causing a computer to execute a learning process,
Performing a pre-processing on a plurality of learning patterns in which a belonging class is designated in advance;
Extracting a feature amount from each of the plurality of learning patterns subjected to the preprocessing;
Classifying each of the plurality of learning patterns into one of a plurality of classes based on a corresponding feature value;
For each of the plurality of types of classes, the degree of scatter within the class indicating the distribution state in the corresponding class for the feature value corresponding to each of the plurality of learning patterns classified into each of the plurality of types of classes. Calculating, and further calculating a degree of inter-class dispersion indicating a distribution state between the plurality of types of classes;
Calculating an evaluation value that satisfies any one or both of the following: an evaluation value that increases as the degree of dispersion within the class decreases; and an evaluation value that increases as the degree of dispersion between classes increases. Let the computer run,
The step of calculating the evaluation value learns at least one of the step of performing the preprocessing, the step of extracting or the step of classifying by feeding back the calculated evaluation value so that the evaluation value becomes high Including the step of
The step of calculating includes a step of changing a weight when calculating the degree of dispersion within the class, as the value of the characteristic amount is far from the range in which the value of the characteristic amount for each of the plurality of types of classes is the most distributed. , Learning program.   A recording medium on which the learning program according to claim 23 is recorded.   The pattern recognition apparatus which recognizes the pattern of a process target image using the image processing algorithm of Claim 10 or Claim 11.   A pattern recognition method for recognizing a pattern of an image to be processed using the image processing algorithm according to claim 21 or 22.

Images