JP2007200246A - Method for evaluating image processing algorithm, method, device, and program for generating same algorithm, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To give a high evaluation to an image processing algorithm having a high processing capability. <P>SOLUTION: Image data of which the positions of extraction object parts like defects are known is subjected to preprocessing of an image processing algorithm being an evaluation object and then is subjected to primary evaluation, and candidates of extraction object parts including non-extraction object parts are extracted from an image after the preprocessing, and feature quantities of these candidates are calculated, and candidates are classified to extraction object parts and non-extraction object parts because positions of extraction object parts are known, and image data is subjected to secondary evaluation by a degree of separation of feature quantities between a group of extraction object parts and a group of non-extraction object parts, and parameters of the image processing algorithm are adjusted on the basis of the evaluation value of secondary evaluation. Rejection processing is performed on the basis of primary evaluation. A primary evaluation criterion is changed on the basis of secondary evaluation. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an image processing algorithm evaluation method, generation method, apparatus, program, and recording medium, and in particular, a method for evaluating the performance of an image processing algorithm for extracting a desired target portion from an image of image data, and the evaluation method The present invention relates to a method for generating an image processing algorithm using a computer, and an apparatus, program and recording medium therefor.

  2. Description of the Related Art In recent years, the application field of image processing has been expanding due to lower prices of image processing devices such as CCD (Charge Coupled Device) and image processing devices such as personal computers. In particular, the application of image processing to defect inspection in the industrial field is widespread, increasing the need for product and device yield, quality improvements and cost reductions by reducing visual inspectors. However, an 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. Therefore, at present, a wide variety of image processing algorithms have been developed for defect inspection in accordance with target products and devices, and the development man-hours increase in proportion to the expansion of application fields and inspection objects. In general, the development and introduction of image processing algorithms consist of design work and adjustment work, both of which involve trial and error that repeats the work of combining image processing modules, changing various parameters, and the like, It takes a lot of man-hours. As described above, the image processing algorithm development cost does not decrease while the price of the image processing device is reduced. Therefore, the image processing algorithm development cost in the defect inspection apparatus is increasing.

  Here, since an image processing algorithm requires a very large number of man-hours for its development and introduction, establishment of an automatic generation method of an image processing algorithm is desired as the image processing application field expands.

  The automatic generation of image processing algorithms creates candidate image processing algorithm groups, applies image data whose pass / fail results are known in advance to these, and applies each of the pass / fail judgment accuracy obtained as a result of the application. The evaluation value of the performance of the image processing algorithm is calculated, and a new candidate image processing algorithm group is repeatedly created based on these evaluation values, and an image processing algorithm with higher evaluation is obtained.

  In general, the amount of computation by a CPU (Central Processing Unit) required to derive a desired computation result according to an image processing algorithm is large, and it is very difficult to evaluate a large number of candidate image processing algorithms using image data. Since a large amount of computation is required by the CPU, it takes a long time to generate an image processing algorithm. Further, when the number of image processing algorithms to be evaluated is reduced, the probability that an image processing algorithm with sufficient accuracy can be generated decreases. Therefore, in automatic generation of an image processing algorithm, it is important to evaluate more candidate image processing algorithms in a certain time.

  Patent Documents 1 and 2 disclose conventional techniques for evaluating an image processing algorithm and generating an image processing algorithm based on the evaluation result.

In Patent Document 1, a non-defective image data sample, a defect (defective) image data sample, and an intermediate grade image data sample having intermediate characteristics are input, and the defect portion included in each input image data sample is input. The worker sets the order. Then, the discrimination ability between the non-defective product and the defective product of the image processing algorithm is evaluated based on the correlation between the feature amount of the defective part calculated after the processing of each image data sample by the image processing algorithm and the order set in advance by the operator. Based on the evaluation result, an appropriate one is selected from a plurality of image processing algorithm candidates.

In Patent Document 2, an operator creates a category image file in which an urban area and a non-urban area are expressed in white and black, respectively, with respect to an original image. Then, a feature vector of each image data sample composed of the original image and the category image file is calculated by a feature extraction method generated by combining various image processing program modules, and the main feature vector of the calculated feature vector is calculated. A primary evaluation value is calculated from a compression feature vector as a component. A feature space is configured by combining a plurality of compressed feature vectors having high calculated primary evaluation values, and a secondary evaluation for evaluating the feature space based on the independence is performed, and the secondary evaluation value is the highest. A tertiary evaluation is performed based on the recognition rate in the feature space.
JP-A-4-367882 JP 2004-38744 A

  In the technique of Patent Document 1, an image processing algorithm to be evaluated is configured by combining an image processing module and a feature extraction module in combination with a final-stage image processing module. Is going. That is, after determining all the processing procedures of the image processing algorithm, the discrimination ability can be evaluated only by performing all of the processing on the learning image data. Therefore, even for an image processing algorithm that reveals that the discrimination capability is insufficient after only a few stages of processing procedures, the processing procedure is carried out up to the final stage, so that a large amount of computation is required.

  In the technique of Patent Document 2, the image processing algorithm is evaluated in three stages: feature extraction method evaluation, feature space evaluation, and recognition rate evaluation. The feature extraction method evaluation unit uses an image processing algorithm having a feature extraction method that is predicted not to exhibit a sufficient recognition rate based on the distribution of feature amounts of parts belonging to each class calculated by the feature extraction method. The amount of computation required for the evaluation of the image processing algorithm is reduced by rejecting the above and not performing the subsequent evaluation.

  However, since statistical processing related to image data conversion processing, feature extraction processing, and feature amount distribution has been performed up to the feature extraction method evaluation unit, the technology disclosed in Patent Document 1 is generally used before the evaluation is rejected. An amount of calculation equal to or greater than that is required, and the amount of calculation has not decreased. In addition, since the evaluation criterion in each evaluation unit is fixed during the generation of the image processing algorithm, it is sufficient if the correlation between the evaluation criterion in the last recognition rate evaluation unit and the evaluation criterion in the previous evaluation unit is not high. The recognition rate may not be obtained. In other words, even with a feature extraction method that has obtained an evaluation value that is not sufficiently high in the feature extraction method evaluation criteria, there may be an image processing algorithm that shows a high evaluation value in the recognition rate evaluation unit depending on the combination of feature extraction methods. However, since these feature extraction methods have already been rejected, the final evaluation is not reached, and the image processing algorithm including these feature extraction methods has a high recognition rate. It is never generated.

  SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide an image processing algorithm evaluation method, apparatus, program, and storage medium capable of evaluating an image processing algorithm according to processing capability.

Another object of the present invention is to provide a method, an apparatus, a program, and a recording medium for generating an image processing algorithm evaluated according to a processing capability.

  Still another object of the present invention is to provide an image processing algorithm evaluation method, generation method, apparatus, program, and recording medium that require a small amount of calculation.

  According to an aspect of the present invention, the image processing algorithm evaluation method converts the image data of the processing target image so that it can be processed, and calculates the plurality of defect candidate portions in the image based on the converted image data An image processing algorithm that uses a quantity to extract one or more true defect portions from an image to be processed is evaluated using image data having predetermined information about one or more true defect portions to be extracted included in the image And has the following characteristics.

  A conversion process for performing conversion processing of an image processing algorithm on image data having predetermined information, converted image data converted so as to be processed by applying the conversion process, and conversion based on the predetermined information A feature amount is calculated by a primary evaluation step for evaluating processing capability, a feature amount calculation step for calculating one or more types of feature amounts for each of a plurality of defect candidate portions of image data, and a feature amount calculation step based on predetermined information. A classification step of classifying the plurality of defect candidate portions into true defect portions and the rest as pseudo defect portions; a group of one or more feature amounts of the true defect portion; and a group of one or more feature amounts of the pseudo defect portion; Based on at least one of a separation degree calculation step for calculating a separation degree in a feature space defined by one or more types of feature amounts, a separation degree calculated by the separation degree calculation step, and an evaluation result of the primary evaluation step. And a second evaluation step of evaluating the performance of the image processing algorithms Te.

  The conversion step performs conversion processing of an image processing algorithm on the image data to be processed, the primary evaluation step evaluates the converted image data based on predetermined information, and the feature amount calculation step includes a plurality of pieces of image data. One or more types of feature amounts for each defect candidate portion are calculated, and the classification step classifies the plurality of defect candidate portions for which the feature amounts are calculated into a true defect portion and a pseudo defect portion based on predetermined information, and a separation degree calculation step Calculates the degree of separation in the feature space between the group of one or more feature quantities of the true defect part and the group of one or more feature quantity of the pseudo defect part. In the secondary evaluation step, the degree of separation and the primary evaluation are calculated. The performance of the image processing algorithm is evaluated based on at least one of the process evaluation results.

  Therefore, it is possible to evaluate the performance of the image processing algorithm according to the conversion processing capability based on the evaluation result of the primary evaluation step, and it is possible to evaluate the performance of the image processing algorithm according to the separation capability based on the evaluation result of the secondary evaluation step. Further, it is possible to evaluate the performance of the image processing algorithm according to the conversion ability and the separation ability based on the evaluation results of both the primary evaluation and the secondary evaluation.

  Preferably, the image processing algorithm evaluation method includes a rejection step of rejecting an image processing algorithm that does not satisfy the evaluation criteria based on the evaluation result of the primary evaluation step.

  Since the rejection process rejects an image processing algorithm that does not satisfy the evaluation criteria based on the evaluation result of the primary evaluation process, the image processing algorithm that is predicted to have no sufficient degree of separation in the secondary evaluation process according to the evaluation criteria is It can be rejected at the time of completion of the primary evaluation process in which only the image data conversion process with a relatively small amount of calculation is performed. Therefore, it is possible to evaluate the image processing algorithm with a smaller amount of calculation.

  Preferably, the image processing algorithm evaluation method includes a primary evaluation criterion changing step of changing an evaluation criterion of the primary evaluation step based on an evaluation result of the secondary evaluation step.

In the primary evaluation standard changing process, since the evaluation standard of the primary evaluation process is changed based on the evaluation result of the secondary evaluation process, the evaluation standard of the primary evaluation process may be made to follow the evaluation standard of the secondary evaluation process. It becomes possible.

  Preferably, the image data having the predetermined information includes primary evaluation reference image data in which one or more true defect portions to be extracted included in the processing target image and other pseudo defect portions are expressed by different density values. Therefore, the primary evaluation reference image data can be easily provided. Furthermore, by using images with different density values, for example, binary images, the size of image data having predetermined information can be reduced, and the amount of data processing during evaluation can be reduced.

  Preferably, in the primary evaluation step, the evaluation is performed by comparing the primary evaluation reference image data with the converted image data.

  Preferably, in the primary evaluation criterion changing step, the primary evaluation criterion image data is changed using the converted image data converted by the image processing algorithm that has been evaluated in the secondary evaluation step. Therefore, it is possible to omit the process of newly creating the primary evaluation reference image data.

  Preferably, in the primary evaluation standard changing step, the post-conversion image data converted by the image processing algorithm that obtains the best evaluation in the secondary evaluation step is replaced with the primary evaluation standard image data. Therefore, since the image conversion processing result data based on the image processing algorithm to which the best secondary evaluation is given is used as the primary evaluation reference image data, the image processing algorithm to which the best evaluation of the separation ability is given is the primary evaluation. It is possible to obtain a high evaluation value even at.

  According to another aspect of the present invention, an image processing algorithm generation method generates a secondary image in addition to the steps of the image processing algorithm evaluation method described above in order to generate an image processing algorithm using the image processing algorithm evaluation method described above. Subsequent to the evaluation step, based on a predetermined criterion, a completion determination step for determining whether or not the generation of the image processing algorithm has been completed, and if the completion determination step determines that the generation has not been completed, the primary evaluation step and Adjustment that creates one or more image processing algorithms by changing the elements of the image processing algorithm based on the evaluation result output by at least one of the secondary evaluation steps, and outputs the created image processing algorithm to the conversion step A process.

  Preferably, the method further includes a rejection step of rejecting an image processing algorithm that does not satisfy the evaluation criteria based on an evaluation result in the primary evaluation step.

  Preferably, in the rejection process, the image processing algorithm is rejected in accordance with the rejection rate determined based on the processing capability after the rejection process, so that the processing in the subsequent process only for the image processing algorithm is not performed.

  Accordingly, it is possible to determine the rejection rate in accordance with the processing capacity after the rejection process, and thereby reduce the number of image processing algorithms to be subjected to secondary evaluation. As a result, the primary evaluation, rejection, secondary evaluation, and processing related to generation of the image processing algorithm constituting one cycle of generation of the image processing algorithm can be performed at high speed. Therefore, more image processing algorithms can be evaluated within a certain time.

  Preferably, in the rejection process, only the image processing algorithm is rejected by rejecting the image processing algorithm according to the rejection rate determined based on the correlation state between the evaluation result of the primary evaluation process and the evaluation result of the secondary evaluation process. Processing in subsequent steps is not performed.

When the correlation between the evaluation result of the primary evaluation process and the evaluation result of the secondary evaluation process is high, the evaluation rank in the primary evaluation and the evaluation rank in the secondary evaluation of a certain image processing algorithm may change. It is thought that it will decrease. Therefore, in this case, a high rejection rate can be determined based on the correlation state between the evaluation result of the primary evaluation process and the evaluation result of the secondary evaluation process, and more image processing algorithms can be evaluated within a certain time. Enable.

  By applying the pipeline configuration to the image processing algorithm generation device by determining the rejection rate so that the two processes of primary evaluation and rejection and secondary evaluation and image processing algorithm generation can be performed in the same processing time In this case, there is no bottleneck at the time of pipeline processing, and high performance can be obtained.

  The rejection rate is dynamically determined according to the correlation state with an upper limit of the rejection rate that allows the first evaluation and rejection processing and the second evaluation and image processing algorithm generation processing to be performed in the same processing time. As a result, more image processing algorithms can be evaluated within a certain period of time.

  Preferably, the image processing algorithm generation apparatus includes two or more conversion means and a series of means including a feature amount calculation means, a classification means, and a separation degree calculation means, and evaluates two or more image processing algorithms in parallel. To do.

  Therefore, a processing apparatus (for example, a DSP) suitable for preprocessing corresponding to the conversion means and a processing apparatus (for example, a RISC system CPU) suitable for post-processing corresponding to the separation degree calculation means from the feature amount calculation means are used. By performing parallel processing, it is possible to configure a processing device according to the required image processing algorithm generation speed.

  The image processing algorithm generation apparatus is not completed by the primary completion determination means for determining whether or not the primary evaluation is completed based on the evaluation result output by the primary evaluation means, and the primary evaluation completion determination means. When it is determined, one or more image processing algorithms are created by changing any element of the image processing algorithm based on the evaluation result output by the primary evaluation unit, and the created image processing algorithm is converted into a conversion unit And a primary adjustment means for outputting to.

  Therefore, processing up to the primary evaluation is performed by a processing device suitable for preprocessing (for example, a DSP), and a new image processing algorithm is generated when the primary evaluation is completed. A secondary evaluation requiring a large amount of computation is performed only for an image processing algorithm that realizes a cycle and has obtained a sufficiently high primary evaluation. As a result, it is possible to evaluate more image processing algorithms with a small amount of calculation.

  In still another aspect of the present invention, there are provided an apparatus corresponding to the above-described image processing algorithm evaluation method, a program for causing a computer to execute the method, and a computer-readable recording medium on which the program is recorded.

  In still another aspect of the present invention, an apparatus corresponding to the above-described image processing algorithm generation method, a program for causing a computer to execute the method, and a computer-readable recording medium on which the program is recorded are provided.

According to the present invention, the image processing algorithm is evaluated in two stages, and an image processing algorithm that is considered not to have sufficient separation capability is rejected after the primary evaluation and before the secondary evaluation. The amount of calculation can be reduced. In general, when the required amount of computation up to the primary evaluation step is compared with the amount of computation required from the primary evaluation step to the secondary evaluation step, the latter amount of computation is larger, so image processing is performed in the primary evaluation. Rejecting the algorithm is desirable from the viewpoint of reducing the amount of calculation.

  Further, according to the present invention, since the evaluation criterion for the primary evaluation of the image processing algorithm is changed based on the primary evaluation result of the image processing algorithm that has been subjected to the secondary evaluation, the interval between the primary evaluation and the secondary evaluation. Correlation can be improved. For this reason, for example, even if an image processing algorithm with a low primary evaluation is rejected, an image processing algorithm that can take a high secondary evaluation proceeds to the steps after the primary evaluation step. A high evaluation value can be given to an image processing algorithm with high ability.

  In particular, when there is an error in the predetermined information that is the evaluation criterion used for the primary evaluation, the evaluation axis of the primary evaluation criterion and the secondary evaluation criterion are different, and thus the evaluation value may not converge when generating the image processing algorithm. There is. In this case, by changing the evaluation reference image data of the primary evaluation based on the converted image data that is the basis of the result of the primary evaluation of the image processing algorithm that gives the best secondary evaluation, the predetermined information The error can be corrected.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
(Embodiment 1)
FIG. 1 shows a processing procedure for image processing algorithm evaluation and image processing algorithm generation using the image processing algorithm evaluation according to the first embodiment, and FIG. 2 shows image processing algorithm evaluation and generation according to the first embodiment. FIG. 3 shows the configuration of a computer on which the device according to each embodiment is mounted.

  Generally, an image processing algorithm is a combination of several image processing modules (binarization processing module, filter processing module, etc.), and each image processing module has some parameters such as binarization threshold and filter size. Have. That is, the image processing algorithm is expressed by a combination of various image processing modules and the parameter values of each module. This image processing algorithm generation method repeats changing the combination of image processing modules and the parameter values of each module based on the evaluation values calculated by the image processing algorithm evaluation method, thereby obtaining an image with a higher evaluation value. A processing algorithm is generated. It is assumed that the image processing module is selected from those registered or generated in advance by an operator or a program.

  Here, the original image data obtained by imaging the product to be processed by the image processing algorithm, and the position information and shape in the original image of the extraction target portion such as a true defect portion of the product in advance with respect to the image data A set with reference image data to which information is given by an operator is called a learning image data set.

  In addition, here, the image processing algorithm includes a pre-processing unit that performs image processing such as noise removal and edge enhancement by filtering on image data to be processed, and a defective portion for an image that has undergone image quality improvement. After calculating the image feature amount of each defect candidate portion (hereinafter referred to as an extraction candidate region), which is a plurality of partial regions that are candidates for extraction, a plurality of defects are determined according to a threshold value set for the feature amount It is assumed that the candidate part is composed of a post-processing part that separates the extraction target part of the true defect part and the non-extraction target part of the pseudo defect part. The pseudo defect portion refers to a portion that is erroneously detected as a side effect when a true defect portion is detected.

(Computer structure and recording medium)
Referring to FIG. 3, the computer includes a mouse 110, a keyboard 350 and a pen / tablet 360, a monitor 310 formed of a CRT (cathode ray tube) or a liquid crystal, and the computer itself. CPU (abbreviation of central processing unit) 312 for centralized control, memory 324 including ROM (read only memory) or RAM (abbreviation of random access memory), fixed disk 326 such as hard disk, FD ( A flexible disk) 332 is detachably mounted, and an FD drive device 330 that accesses the mounted FD 332 and a CD-ROM (Compact Disc Read Only Memory) 342 are detachably mounted, and the mounted CD-ROM 342 is removed. C to access
A D-ROM drive 340 and a communication interface 380 for connecting the computer to various wired or wireless communication networks 109 including the Internet are included. These units are connected for communication via a bus. Various types of information can be transmitted / received to / from the outside via the communication network 109. The received information includes image data to be processed, and the image data may be data of an image taken by an external camera, for example.

  The computer shown in FIG. 3 may be provided with a magnetic tape device in which a cassette type magnetic tape is detachably mounted to access the magnetic tape.

  The memory 324 is provided with a learning image data storage unit 34, and stores converted image data 13, primary evaluation reference image data 12, image processing algorithm data 14, and algorithm generation / adjustment data 44.

  As shown in FIG. 2, the learning image data storage unit 34 stores a plurality of learning image data sets 35. The learning image data set 35 includes original image data 36 to be processed by the image processing algorithm and reference image data 37 in association with each other. The reference image data 37 includes position information 38 indicating the position (coordinates) and shape information 39 indicating the shape of the extraction target part as predetermined information for specifying each extraction target part in the original image of the related original image data 36. In addition, image data in which one or more extraction target portions included in the original image and other non-extraction target portions are expressed by different density values is included. An image data input unit 1 described later acquires original image data 36 including a defect to be extracted by imaging or the like, and stores it in the learning image data storage unit 34. At that time, an extraction target information setting unit 21 described later is for an operator to input the position information 38 and the shape information 39 of the extraction target unit to the original image data 36 using an external input device such as a keyboard 350. Provide functionality. The reference image data 37 is constituted by the position information 38 and the shape information 39 input here, and a learning image data set 35 composed of the reference image data 37 and the original image data 36 is a learning image data storage unit. 34.

  As shown in FIG. 2, the image processing algorithm data 14 includes values of one or more image processing modules 42 and one or more image processing parameters constituting an image processing algorithm having the best evaluation (highest evaluation). 43, and an evaluation value 41 of the image processing algorithm. Although the parameter value 43 is assumed here, the image processing algorithm data 14 may not include the parameter value 43 because the parameter may not be included depending on the image processing algorithm.

  The algorithm generation / adjustment data 44 is referred to in order to generate or adjust an image processing algorithm to be processed, and a plurality of image processing modules 42 that can be components of the image processing algorithm and a plurality of parameter values for image processing. 43. The image processing module 42 is expressed by a parameter value 43 cited therein.

Referring to FIG. 2, the apparatus includes an image data input unit 1, a display unit 3 corresponding to a monitor 310 for displaying various types of information to an operator, and an image processing algorithm for connecting the image data input unit 1 and the display unit 3. A generation unit 2 is provided. The image processing algorithm generation unit 2 corresponds to an image processing device. The image processing algorithm generation unit 2 includes an extraction target information setting unit 21 that refers to the learning image data storage unit 34, an image processing algorithm generation / adjustment unit 22 that refers to the algorithm generation / adjustment data 44, an image data conversion unit 23, A primary evaluation unit 24 that evaluates the image data conversion performance of the image processing algorithm based on the primary evaluation reference image data 12, and a rejection determination unit 25 that determines whether or not to reject the image processing algorithm based on the result of the primary evaluation. The converted image data storage processing unit 26, the feature amount calculation unit 27, the classification unit 28, the separation degree calculation unit 29, and the secondary evaluation unit that store the image data 13 converted by the image data conversion unit 23 in the memory 324. 30. Image processing algorithm for storing the image processing algorithm evaluated as the best in the memory 324 as the image processing algorithm 14 Storage processing unit 31, and a primary criterion image data 12 primary evaluation criterion changing unit 32 rewrites the and image processing algorithm generating completion determining unit 33 for determining the completion of the generation of the image processing algorithms. Each of these units is prepared in advance as a program that can be executed by the computer of FIG. 3, and the program is stored in advance in the memory 324 or the like, and each program is read and executed by the CPU 312.

  The evaluation procedure, generation procedure, and apparatus of the image processing algorithm according to the present embodiment will be outlined with reference to FIGS. 1 and 2, and then the operation using the actual learning image data set 35 will be described with reference to the drawings. .

  In FIG. 1, a primary evaluation is performed on pre-processed image data with respect to an arbitrary image processing algorithm, and the separation ability between the extraction target portion and the non-extraction target portion after post-processing is evaluated. A secondary evaluation is performed. The flowchart of FIG. 1 shows an image processing algorithm evaluation method and an image processing algorithm generation method using the image processing algorithm evaluation method. The parts corresponding to the image processing algorithm evaluation method are STEP2, STEP3 to STEP10, and STEP12.

  In STEP 1, the image processing algorithm generation / adjustment unit 22 generates an image processing algorithm to be evaluated by combining one or more modules 42 or one or more parameter values 43 read from the algorithm generation / adjustment data 44. To do. The type of the module 42 or the type of the parameter value 43 to be read out from the algorithm generation / adjustment data 44 and used for the generation is determined by the operator operating an external input device such as the keyboard 350. Is directed against.

  In STEP 2, the extraction target information setting unit 21 selects a learning image data set 35 used for image processing algorithm evaluation from the learning image data storage unit 34 based on a selection signal input by the operator operating the keyboard 350. Read selectively.

  In STEP 3, the image data conversion unit 23 applies the preprocessing unit of the image processing algorithm to be evaluated to the original image data 36 of the learning image data set 35 that has been selectively read out, and converts the converted image data 13 is generated.

  In STEP 4, the primary evaluation unit 24 processes the converted image data 13 so that the converted image data 13 can be compared with the image data indicated by the primary evaluation reference image data 12. Compared with the image data indicated by the next evaluation reference image data 12, a primary evaluation of the image processing algorithm to be evaluated is performed based on the comparison result. Hereinafter, the evaluation value of the primary evaluation is referred to as a primary evaluation value.

  Here, the primary evaluation reference image data 12 can be the reference image data 37 of the learning image data set 35 selected in STEP 1. The primary evaluation reference image data 12 can be set separately by the operator.

In STEP 5, the rejection determination unit 25 determines whether the primary evaluation value output by the primary evaluation unit 24 is equal to or greater than a predetermined reference value stored in advance in the memory 324 or the like. If it is determined that the value is less than the predetermined reference value, the image processing algorithm is rejected. That is, the image processing algorithm whose primary evaluation value is smaller than the predetermined reference value is controlled so that subsequent processing, that is, processing after STEP 6 is not performed. If it is determined that the value is equal to or greater than the predetermined reference value, the subsequent processing is executed for the image processing algorithm.

  In STEP 6, the post-conversion image data storage processing unit 26 stores the post-conversion image data 13 in the memory 324 for an image processing algorithm whose primary evaluation value is greater than or equal to a predetermined reference. The converted image data 13 stored in the memory 324 is used when the primary evaluation reference image data 12 is changed later.

  In STEP 7, the feature amount calculation unit 27 determines a plurality of extraction candidate regions in the image indicated by the post-conversion image data 13, and calculates position information and image feature amounts for each. Here, determination of extraction candidate regions and their position information can be acquired by binarizing the converted image data 13. However, the procedure for determining the extraction candidate region and acquiring the position information is not limited to this. Based on the extraction candidate area and the position information obtained in this way, the extraction candidate area in the image of the converted image data 13 is determined, and the image feature amount of the same area is calculated.

  In STEP8, the classification unit 28 compares the position information 38 of the reference image data 37 of the learning image data set 35 selected by the operator with the position information acquired in STEP7, and is acquired in STEP7 based on the comparison result. It is determined whether each extracted candidate region is an extraction target portion or a non-extraction target portion.

  In STEP 9, the degree-of-separation calculating unit 29 sets a feature space configured based on the image feature amount calculated in STEP 7. Then, the calculated feature amount of each extraction target area that is an extraction candidate area determined as an extraction target area in STEP 8 and each non-extraction target area that is an extraction candidate area determined as a non-extraction target area are calculated. Each feature amount is plotted in the aforementioned feature space. Then, the degree of separation indicating the degree of separation between the feature amount group plotted for the extraction target region and the feature amount group plotted for the non-extraction target region in the feature space is calculated.

  In STEP 10, the secondary evaluation unit 30 performs secondary evaluation of the image processing algorithm to be evaluated based on the degree of separation calculated in STEP 9. Hereinafter, the evaluation value of the secondary evaluation is referred to as a secondary evaluation value. Details of the secondary evaluation will be described later.

  In STEP 11, the image processing algorithm storage processing unit 31 stores the image processing algorithm and the secondary evaluation value in the memory 324 as the image processing algorithm data 14 based on the secondary evaluation value.

  In STEP 12, the primary evaluation criterion changing unit 32 changes (rewrites) the primary evaluation criterion image data 12 in the memory 324 based on the evaluation result of the image processing algorithm. As an example, for the image processing algorithm in which a secondary evaluation value indicating high evaluation is output, the image data 13 after conversion by the image processing algorithm is already stored in the memory 324 in STEP 4. It is possible to change the primary evaluation reference image data 12 indicated by the reference image data 37 of the learning image data set 35 previously selected by the operator. For example, the primary evaluation reference image data 12 is changed by replacing the primary evaluation reference image data 12 with the converted image data 13.

  Next, an image processing algorithm generation method using the above-described image processing algorithm evaluation method will be described with reference to FIG.

  The processing of STEP1 to STEP10 is performed, and the secondary evaluation value of the image processing algorithm is calculated. In STEP11, every time the processing of STEP1 to STEP10 is repeated, the image processing algorithm having the largest secondary evaluation value among the evaluated image processing algorithms is set as the best image processing algorithm, and the secondary evaluation value, the image processing module, and the parameter are set. The combination of values is stored in the memory 324 as the image processing algorithm data 14.

  Thereafter, every time it is repeated, the image processing algorithm storage processing unit 31 stores the memory only when the secondary evaluation value is an image processing algorithm that exceeds the evaluation value 41 of the image processing algorithm data 14 stored in the memory 324. The image processing algorithm data 14 of 324 is rewritten using the data of the image processing algorithm.

  In STEP 13, the image processing algorithm generation completion determination unit 33 determines whether an end condition is satisfied. If the termination condition is not satisfied (NO in STEP 13), a new module 42 or parameter value 43 adjusted by the image processing algorithm generation / adjustment unit 22 using the algorithm generation / adjustment data 44 in STEP 14 is adjusted. Generate an image processing algorithm. Then, the processing after STEP 3 is repeated in the same manner for the newly generated image processing algorithm.

  In this way, when it is determined that the end condition is satisfied while the generation of a plurality of image processing algorithms is repeated (YES in STEP 13), the generation of the image processing algorithm based on the evaluation value ends. Therefore, at this time, the image processing algorithm indicated by the image processing algorithm data 14 stored in the memory 324 has the largest secondary evaluation value among the image processing algorithms generated so far, that is, the best image processing algorithm. Become.

  Note that the best rewriting standard of the image processing algorithm data 14 in STEP 11 is not limited to the above, and a plurality of image processing algorithm data 14 may be stored according to the data capacity that can be stored in the memory 324. Is possible. For example, the image processing algorithm data 14 of the image processing algorithm having the second largest evaluation value from the top to the third is stored, and an image processing algorithm exceeding the evaluation value from the top to the third is generated. The image processing algorithm data 14 in the memory 324 may be rewritten. In this way, a plurality of different image processing algorithms having the same secondary evaluation value can be acquired.

  Here, the termination condition of STEP 13 will be described. The end condition is, for example, when a value exceeding the reference value of the secondary evaluation value set in advance by an operator or a program is generated, or the parameter adjustment of the image processing algorithm in STEP 14 and the re-evaluation of the image processing algorithm are repeated. For example, when the number of times exceeds a specified number of times set by an operator or a program, or when the total processing time exceeds a set time.

  If the end condition is not satisfied, the image processing algorithm is adjusted in STEP14. Specifically, the combination of the image processing modules 42 and the parameter value 43 are changed based on the evaluation value of the image processing algorithm. As a technique for optimizing the combination of the image processing modules 42 and the parameter value 43 based on a certain evaluation value, there are many proposals such as those using GA (genetic algorithm) and SA (simulated annealing). Has been. In either case, for combinations and parameter values with high evaluation values, those combinations and values are slightly changed, and a method with a higher evaluation value is searched.

Regarding the change of the “combination of image processing modules 42”, for example, JP-A-2002-366929 discloses a method using GA (genetic algorithm) as a prior art document. By applying the method disclosed in this publication, it is possible to optimize the combination of image processing modules and the parameter value with respect to the evaluation value.

  Here, an expression processing method and a change method (mutation and crossover) of the image processing module group related to GA will be described with reference to FIGS. 4 (A), 4 (B), and 4 (C).

<Expression method of image processing module group>
The image processing module group is formed by connecting a plurality of image processing modules, and the image processing module has a parameter value. In FIG. 4A, the image processing module group is shown in a state where a plurality of image processing modules indicated by ellipses are connected. In FIG. 4A, for example, for the character string “F1-1” in the first ellipse, “F1” indicates the type of the image processing module of the ellipse, and “1” next to the hyphen (“−”). Represents a parameter value of the image processing module. The image processing module types and parameter values are similarly expressed for the other elliptical image processing modules.

<Mutation>
The mutation in GA is to expect that an image processing module group with higher evaluation (large evaluation value) is generated by randomly changing image processing modules or parameters of the image processing module group. is there. FIG. 4B shows a state in which the image processing module group B2 is generated by the mutation that the second image processing module and the parameter value of the image processing module group B1 are changed.

<Crossover>
The crossover in the GA means that the image processing modules are exchanged from the middle of the connection for two image processing module groups. Usually, the image processing module groups with high evaluation are crossed with the expectation that an image processing module group with higher evaluation (high evaluation value) is generated. FIG. 4C shows a state in which crossover is performed on the image processing module groups B1 and B3. Here, by replacing the partial module group BB1 in the middle of the connection of the image processing module group B1 and the partial module group BB3 in the middle of the connection of the image processing module group B3, the image processing module group B1 is replaced with the image processing module group B11. The image processing module group B3 is changed to the image processing module group B31.

  If the number of combinations of the image processing modules 42 and the range taken by the parameter value 43 are sufficiently small regarding the satisfaction of the termination condition of STEP 13, all combinations of the image processing modules 42 can be performed without waiting for the termination condition of STEP 13 to be satisfied. It is possible to finish the evaluation for all image processing algorithms that can be generated under the condition where the parameter value 43 is changed. In this manner, the image processing algorithm having the best evaluation can be obtained by repeatedly applying the procedure after STEP 3 to the image processing algorithm generated every time the combination of the image processing modules and the parameter values are changed in STEP 14. Automatic generation of algorithms can be realized.

Next, application of the image processing algorithm evaluation method to the actual learning image data set 35 will be schematically described with reference to FIG. The original image data 36 of the learning image data set 35 is subjected to filter processing STEP3a, which is a preceding process corresponding to STEP3 in FIG. 1, by an image processing algorithm to be evaluated, and is converted into filtered image data 131. In order to compare the post-filter image data 131 with the primary evaluation reference image data 12 that is binarized image data, a binarization process STEP3b that is a subsequent process corresponding to STEP3 is performed, and binarization conversion is performed. The post image data 132 is obtained. Hereinafter, the binarized image data 132 is referred to as converted image data 13.

  In the filter processing STEP3a, the edge is emphasized for the processing target image, the dilation processing is repeated a plurality of times for the image with the edge emphasized, and then the shrinkage processing is repeated a plurality of times.

  Here, the primary evaluation reference image data 12 uses the reference image data 37 as it is. However, the primary evaluation reference image data 12 is not limited to this.

  Further, regarding the threshold value for binarizing the filtered image data 131, the number of extraction objects (the number of position information 38) is designated in advance by the reference image data 37. Therefore, extraction according to the number is performed. It is possible to determine the binarization threshold value so that the candidate area appears and input it from the keyboard 350 or the like. In the present embodiment, since it is known in advance that the reference image data 37 includes two extraction targets, at least two or more extraction candidate regions appear, and a large amount of calculation is required in the subsequent processing. In order to avoid this, a binarization threshold value is used so that five extraction candidate regions appear. However, the process for causing the extraction candidate region to appear is not limited to the present embodiment, and various methods can be selected as the binarization threshold value determination method.

  Next, a primary evaluation value is calculated by comparing the converted image data 13 (binarized image data 132) with the primary evaluation reference image data 12. For example, the absolute value difference image data between the converted image data 13 and the primary evaluation reference image data 12 is obtained, and the primary evaluation value is defined according to the reciprocal of the density average value of the image obtained from the absolute value difference image data. Is possible. Here, since the comparison is between binary images (generally, difference processing), it is possible to compare at high speed.

  Here, the absolute value difference image is an image indicated by a luminance difference for each pixel between the converted image data 13 and the primary evaluation reference image data 12. The average density value of the absolute value difference image is an average value calculated from the density histogram of the absolute value difference image. Specifically, for example, when the absolute value difference image A is composed of 9900 pixels having a density of 0 (black) and 100 pixels having a density of 255 (white), the average density value of the absolute value difference image A = ( 0 * 9900 + 255 * 100) / (9900 + 100) = 2.55. For example, when the absolute value difference image B is composed of 7000 pixels having a density of 0 (black) and 3000 pixels having a density of 255 (white), the density average value of the absolute value difference image B = (0 × 7000 + 255 × 3000). ) / (7000 + 3000) = 76.5.

  The reciprocal of the calculated average density value defines the primary evaluation value. That is, in the case of the density average value 2.55 of the absolute value difference image A described above, the reciprocal thereof is (1 / 2.55) = 0.392, and in the case of the density average value 76.5 of the absolute value difference image B, The reciprocal is (1 / 76.5) = 0.130. That is, the fact that the reciprocal of the average density value (primary evaluation value) is large means that the difference between the converted image data 13 (binarized image data 132) and the primary evaluation reference image data 12 in FIG. This means that the applied image processing algorithm has an excellent function (has high ability) to bring the image of the original image data 36 closer to the image of the primary evaluation reference image data 12.

  Here, when the average density value of the absolute value difference image data becomes 0, the primary evaluation value is maximized. Further, by weighting the extraction target part according to the importance, when a more important extraction target part appears as an extraction candidate area, a larger evaluation value (high evaluation) appears conversely as the extraction candidate area. If not, it is possible to give a smaller evaluation value (low evaluation).

  More specifically, in STEP4 of the primary evaluation, weighting is performed by adding “weighted image” data stored in advance in the memory 324 or the like to the absolute value difference image data in units of pixels. By the weighting, for example, the primary evaluation value is lower when the extraction target part is not extracted than when the non-extraction target part is extracted if the parts have the same area. It is possible to give a high evaluation (large evaluation value) to the image processing algorithm that can be extracted.

  Note that the method of defining the primary evaluation value is not limited to the present embodiment, and a comparison with a multi-value image is also possible.

  Based on the primary evaluation value thus obtained, it is determined in STEP 5 of FIG. 1 whether or not the subsequent processing is to be performed regarding the image processing algorithm to be evaluated. If the primary evaluation value is greater than or equal to the reference value, the process proceeds to the next process. That is, regarding the post-filter conversion image data 131, the extracted candidate region determined post-conversion image data 133 is obtained by enclosing a portion including the extraction candidate region appearing in the binarized post-conversion image data 132 with a rectangle. Here, an image feature amount is calculated for an extraction candidate region on the image of the image data 133. This corresponds to the processing of STEP 7 in FIG.

  As an example of the image feature amount, there are a contrast defined by a difference in image density between an extraction candidate region and a neighboring region, an area of a region appearing after binarization processing, and the like. Further, based on the primary evaluation reference image data 12, the extraction candidate areas of the converted post-conversion image data 133 are classified into extraction target areas 135 and non-extraction target areas 136. An image of the extracted image classified post-conversion image data 135 is obtained by enclosing a portion including a bold line rectangle and surrounding a portion including the non-extraction target region 136 with a broken line rectangle. This corresponds to the processing of STEP8 in FIG.

  The image feature amount calculated for the extraction candidate region that is the extraction target region 135 and the image feature amount calculated for the extraction candidate region that is the non-extraction target region 136 are arranged in the feature space 137. The feature space 137 corresponds to a two-dimensional coordinate space defined by orthogonal coordinate axes of the feature amount CH1 and the feature amount CH2, assuming that two types of feature amounts CH1 and CH2 are calculated. The feature amount of the extraction target region 135 and the feature amount of the non-extraction target region 136 are plotted. The feature value of the extraction target area 135 is plotted with a thick solid circle (circle), and the feature value of the non-extraction target area 136 is plotted with a thin dotted line (circle).

  The degree of separation (separation degree S) is defined by the distribution state of the plotted feature values of the extraction target region 135 and the plotted feature values of the non-extraction target region 136 in the feature space 137.

  Here, in order to obtain the degree of separation S, a coordinate space (coordinate plane) defined by two coordinate axes defined by two types of feature amounts CH1 and CH2 is used. The feature amount of the region was plotted on a two-dimensional feature space (plane) spanned by a two-dimensional vector (plane vector). However, the feature space to be applied is not limited to the coordinate plane. For example, it may be defined on a one-dimensional coordinate axis defined by one type of feature quantity, or may be defined in a coordinate space defined by each coordinate axis of d (≧ 3) feature quantities.

A coordinate space defined by each coordinate axis of d (≧ 3) feature quantities is a feature space spanned by a d-dimensional space vector. Now, if d feature amounts CHj (j = 1, 2, 3,..., D) are calculated and the calculated d feature amounts are used, the extraction target region and the non-extraction target region are d. Plotted on the feature space spanned by the dimension feature vector CH. In addition,
Feature vector CH = (CH1, CH2,..., CHj,... CHd) ^t (t indicates transposition). On the d-dimensional feature space, the distribution range of points (features) plotted at the positions determined by the feature vector CH for the extraction target region and the points (features) plotted at the positions determined by the feature vector CH for the non-extraction target region The degree of separation S can be obtained according to the distance from the distribution area of (quantity). Therefore, the higher the degree of separation S, the better the ability to separate the extraction target part and the non-extraction target part, that is, the evaluation becomes higher.

  When the non-extraction target region is not extracted depending on the learning image data set 35, the feature amount of the virtual non-extraction target region is plotted at arbitrary coordinates in the feature space 137.

  For example, it is possible that the feature amount of the virtual non-extraction target region is plotted at the origin of the two-dimensional coordinates of the feature space 137.

  Further, for example, the feature amount of the virtual non-extraction target region is a point infinitely far from the origin of the two-dimensional coordinate (a point of the two-dimensional coordinate value defined by the infinity of the feature amount CH1 and the infinity of the feature amount CH2). It can also be plotted.

  In addition, for example, it is possible to plot the feature quantity of the virtual non-extraction target area at a point near the distribution area of the feature quantity of the extraction target area of the two-dimensional coordinates. Here, the neighborhood point is a point inside the concentric ellipse centered on the median of the coordinates where the feature quantities of the plurality of extraction target regions of the two-dimensional coordinates are plotted, for example, at least on or outside the concentric ellipse. It is assumed that the feature amount of one extraction target area is included.

  The calculation of the degree of separation S corresponds to the processing of STEP9. Here, the degree of separation S indicates that the feature quantity distribution area of the extraction target area in the feature space is separated from the distribution area of the feature quantity of the non-extraction target area, and the feature quantity is concentrated in one place in each distribution area. In this case, the degree of separation S increases. Here, when the degree of separation S of the image processing algorithm to be evaluated is sufficiently high, the secondary evaluation value becomes large (secondary evaluation becomes high), and the post-conversion image data 13 processed by the image processing algorithm. Thus, the primary evaluation reference image 12 is replaced. This corresponds to STEP12 in FIG.

  As described above, in this embodiment, since the evaluation criterion for the primary evaluation of the image processing algorithm is changed based on the primary evaluation result of the image processing algorithm having a high secondary evaluation value, Since the correlation between the secondary evaluations increases, even if the image processing algorithm with a low primary evaluation value is rejected, the image processing algorithm that can take a high evaluation value in the secondary evaluation proceeds to the next step. It is possible to give a higher evaluation value to an image processing algorithm with a high separation capability.

In particular, as shown in FIG. 5, when the primary evaluation reference image data 12 used for the primary evaluation includes an error in the designated shape of the extraction target portion, the evaluation value may not converge. In other words, the evaluation of the image processing algorithm is repeated in the process of generating the image processing algorithm, but the primary evaluation reference (primary evaluation reference data 12) and the secondary evaluation reference (feature amount of the feature space 137) are different. That is, since the evaluation axes are different, the evaluation value may not increase (do not converge). At this time, the binarized post-conversion image data 132 of FIG. 5 is changed by changing the standard of the primary evaluation based on the primary evaluation result of the image processing algorithm having a large secondary evaluation value (high secondary evaluation). As shown in Fig. 2, the specified shape error in the extraction target part can be corrected, so that an image processing algorithm that achieves a higher degree of separation can obtain a higher evaluation, and an image processing algorithm that achieves a higher degree of separation. Can be generated. In the image processing algorithm evaluation method according to the present embodiment, the ability of the image processing algorithm to be evaluated is finally defined by the degree of separation between the extraction target region and the non-extraction target region. As shown in the post-conversion image data 132, there is no problem even if a non-extraction target region appears during the primary evaluation.

  In the present embodiment, it is possible to give a higher performance evaluation for an image processing algorithm that can obtain a high degree of separation S in any evaluation step. Further, an image processing algorithm having a low evaluation in the primary evaluation does not obtain a high evaluation in the secondary evaluation, or an image processing algorithm having a high evaluation in the primary evaluation does not obtain a low evaluation in the secondary evaluation. Therefore, even when the image processing algorithm is rejected based on the result of the primary evaluation, an image processing algorithm that can obtain a high degree of separation S can obtain a high evaluation in the secondary evaluation, and the rejection rate can be further increased. Since it can be set sufficiently high, the amount of calculation for image processing algorithm evaluation and generation by the CPU 312 can be reduced.

(Embodiment 2)
When generating an image processing algorithm as an iterative optimization problem, the following method is generally used. In other words, multiple image processing algorithms to be evaluated at one time are created, all or any number of image processing algorithms are evaluated, and image processing is performed based on at least one image processing algorithm with a high evaluation value. By changing the combination of the image processing modules that are constituent elements of the algorithm and the parameter values, the creation of a plurality of image processing algorithms to be evaluated again is repeated until the completion criterion for generating the image processing algorithm is satisfied.

  FIG. 6 is a process flowchart for generating an image processing algorithm by evaluating a plurality of image processing algorithms to be evaluated. The contents of the process having the same reference numerals as those in FIG. 1 in the process flow shown in FIG. 6 are almost the same as those described in FIG. Here, a part different from the processing contents shown in FIG. 1 will be described.

  In STEP 1a of FIG. 6, a plurality of image processing algorithms to be evaluated are newly created by referring to the algorithm generation / adjustment data 44 by the image processing algorithm generation / adjustment unit 22 as in STEP 1 described above.

  In STEP 2, one learning image data set 35 is selected from the learning image data storage unit 34. Although one is selected here, it is also possible to select a plurality. When a plurality of selections are made, the selected learning image data set 35 is applied to each of the plurality of image processing algorithms.

  For each of the created image processing algorithms, the routine Ri (i = 1, 2, 3,..., N) is executed in parallel by commonly applying the learning image data set 35 selected in STEP2. . The routine Ri includes the processes of STEPs 3 to 10 described above.

  As a form of parallel processing, for example, one processor processes a plurality of routines Ri in a time-sharing manner, a plurality of processors individually process the routines Ri in parallel, or a combination of both. Etc. However, the mounting method is not limited to this.

In the present embodiment, it is possible to update the rejection rate by dynamically changing the rejection condition (predetermined reference value) of the image processing algorithm in STEP5. As an example, when the degree of correlation between the primary evaluation value and the secondary evaluation value is low for the image processing algorithm, there is a possibility that a high evaluation value is obtained in the secondary evaluation even if the image processing algorithm has a low primary evaluation. Therefore, the rejection rate must be set low. On the other hand, when the degree of correlation between the primary evaluation value and the secondary evaluation value is sufficiently high, an image processing algorithm with a low primary evaluation obtains a low evaluation in the secondary evaluation with a high probability. It is possible to set, and in this case, it is possible to sufficiently reduce the amount of calculation required.

  After the execution of each routine Ri, in STEP 11 of FIG. 6, the image processing algorithm that is subjected to the secondary evaluation based on the secondary evaluation value output by each routine Ri has a high secondary evaluation. One or a plurality of image processing algorithms are stored.

  With reference to FIG. 7, the case where the procedure of FIG. 6 is applied to the actual learning image data set 35 will be schematically described. In FIG. 7, three evaluations are performed: T-th evaluation → T + 1-th evaluation → T + second-time evaluation. In STEP 1a, first, three image processing algorithms RC, RD, and RE to be evaluated are newly created, and these image processing algorithms are processed in the Tth evaluation on the left side of FIG.

  Next, based on the T-th evaluation result, the image processing algorithms RG, RH, and RI to be newly evaluated by the image algorithm generation / adjustment unit 22 are converted into the image processing module 42 of the algorithm generation / adjustment data 44. And the parameter value 43 are used for the evaluation. This corresponds to the T + 1th evaluation at the center of FIG.

  In FIG. 7, the original image data 36 of the learning image data set 35 selected in STEP 2 is shared for each evaluation. In the T-th evaluation, the reference image data 37 of the learning image data set 35 is used as the primary evaluation reference image data 12.

  In the T-th evaluation, the binarized and converted image data 132 (converted image data 13) by the image processing algorithms RC, RD, and RE is output in each routine Ri. In STEP 4 of each routine Ri of the T-th evaluation, the primary evaluation reference image data 12 and the binarized converted image data 132 are compared, and as a result, primary values 3.0, 1.0, and 2.0 are respectively obtained. An evaluation value is obtained.

  Here, it is assumed that the predetermined reference value for the primary evaluation is set to 1.9. Of the obtained primary evaluation values, the primary evaluation value 1.0 of the image processing algorithm RD is smaller than the predetermined reference value, so that the image processing algorithm RD is excluded from the evaluation target (rejected) at this time, and thereafter Evaluation processing including image processing by the image processing algorithm RD is not performed.

  The image processing algorithms RC and RE are subjected to secondary evaluation in STEP 10 using the feature space 137, and as a result, secondary evaluation values of 1.0 and 5.0 are output, respectively. Of these, the image processing algorithm RE having the largest secondary evaluation value (= 5.0) is determined to be the best, and after the T-th evaluation, the image processing algorithm storage processing unit 31 sets the image processing algorithm data 14 as STEP 14. Stored in the memory 324. At this time, in STEP 11, the primary evaluation criterion changing unit 32 uses the converted image data 13 stored in the memory 324 in STEP 6 of the processing (routine Ri) of the image processing algorithm RE determined to be the best, as the primary evaluation criterion image. Data 12 is replaced.

  Thereafter, the image processing algorithm generation completion determination unit 33 determines in STEP 13 that the generation of the image processing algorithm has not been completed. Next, the T + 1th evaluation is performed.

Prior to the T + 1th evaluation, first, the image processing algorithm generation / adjustment unit 22 refers to the algorithm generation / adjustment data 44 in STEP 14a, and the image processing algorithm RE determined to be the best in the Tth evaluation is used. Then, image processing algorithms RG, RH and RI to be evaluated are generated. These image processing algorithms are newly created by the combination of the image processing modules 42 constituting the image processing algorithm RE that was the best in the T-th evaluation and the change of the parameter value 43. Note that the image processing algorithm RG is assumed to be equivalent to the image processing algorithm RE.

  In the T + 1th evaluation, routines Ri of the image processing algorithms RG, RH, and RI are executed in parallel, and the binarized and converted image data 132 (converted image data 13) is output individually in each routine Ri. These primary evaluation values are assumed to be ∞, 5.0, and 1.5, respectively. Here, since the image processing algorithm RG is equivalent to the image processing algorithm RE that is the best in the T-th evaluation, the maximum value is given to the primary evaluation value. As a result of the primary evaluation, since the primary evaluation value of the image processing algorithm RI is 1.5 and smaller than that of the other image processing algorithms, the subsequent processing of the image processing algorithm RI is not performed (image processing algorithm RI is excluded from evaluation (rejected)).

  For the image processing algorithms RG and RH, it is assumed that secondary evaluation values of 5.0 and 6.0 are given, respectively. At this time, the image processing algorithm RH given the maximum secondary evaluation value 6.0 is stored in the memory 324 as the image processing algorithm data 14 instead of the image processing algorithm RE that was the best in the T-th evaluation. Is done. Further, the primary evaluation reference image data 12 is replaced by the converted image data 13 generated by the processing of the image processing algorithm RH determined to be the best. As a result, the primary evaluation reference image data 12 for the next T + second evaluation indicates the converted image data 13 generated by the processing of the image processing algorithm RH.

  By repeating these series of evaluations, an image processing algorithm having a higher secondary evaluation value can be generated.

(Embodiment 3)
Next, a functional configuration for efficiently realizing the processes of FIGS. 6 and 6 will be described with reference to FIGS. 8 and 9. In the present invention, the image processing algorithm is evaluated in two stages. In the first stage, pre-processing is performed after a pre-processing such as a combination of fixed filter processes, etc., and in the next stage, design is required for each application development destination. Secondary evaluation is performed after complicated post-processing such as feature amount extraction of an image area. For this reason, especially with regard to the hardware configuration for implementing this method, the design up to the primary evaluation is designed by a DSP (digital signal processor) capable of high-speed filtering, and the versatility from the primary evaluation to the secondary evaluation. By designing with an excellent RISC (Reduced Instruction Set Computer) processor, efficient processing becomes possible.

  FIG. 8 schematically shows a functional configuration of an apparatus applied to the present embodiment. In FIG. 8, an image data input unit 1, a RISC image processing apparatus 2A for creating post-processing, secondary evaluation and an image processing algorithm to be evaluated, a display unit 3 for presenting various types of information to an operator, It is mainly composed of a DSP image processing apparatus 4 that performs preprocessing and primary evaluation.

  The DSP image processing apparatus 4 includes an image data conversion unit 23, a primary evaluation unit 24, a storage unit 12a that is a memory for storing the primary evaluation reference image data 12, a rejection determination unit 25, a converted image data storage processing unit 26, and The storage unit 13a is a memory for storing one or more converted image data 13. The DSP image processing apparatus 4 receives the image processing algorithm to be evaluated from the image processing algorithm generation / adjustment unit 22 in the RISC image processing apparatus 2A by a communication function (not shown), and performs the primary evaluation of these image processing algorithms. Only for those whose values satisfy the criteria, the converted image data 13 is transmitted to the feature amount calculation unit 27 of the RISC image processing apparatus 2A.

  A processing procedure for evaluating and generating an image processing algorithm using the functional configuration of FIG. 8 will be described with reference to FIG.

  In FIG. 9, the processes denoted by the same reference numerals as those in FIG. 1 or FIG. 6 are the same as the contents described in FIG. In FIG. 9, a primary evaluation routine Sj (j = 1, 2, 3,..., N) composed of STEP 3 and STEP 4 is executed in parallel for each of the plurality of image processing algorithms generated in STEP 1a or STEP 14a. Further, a secondary evaluation routine Tk (k = 1, 2, 3,..., M) consisting of STEP7 to STEP10 is executed in parallel for each of a plurality of image processing algorithms based on the result of the primary evaluation.

  In the processing, after the primary evaluation by the routine Sj, in STEP 5c, of the N image processing algorithms subjected to the primary evaluation in each routine Sj, for example, the image processing algorithm having the smallest primary evaluation value is rejected. The image processing algorithm is rejected in accordance with a preset rejection condition or a rejection rate of rejecting n of N image processing algorithms having a smaller primary evaluation value. The converted image data 13 of the image processing algorithm having a value is stored. Here, the rejection condition or rejection rate of the rejection process performed in STEP 5c is desired by the operator operating the keyboard 350 by the processing capability of the routine Sj including the pre-processing and the processing capability of the routine Tk including the post-processing. It is possible to set as follows.

  The environment for executing the processing of FIG. 9 is not limited to the functional configuration of FIG. 8. For example, the routine Sj or Tk may be executed in a time-sharing manner by one processor or the like. May be executed in parallel, or may be executed in combination.

  In the present embodiment, it is also possible to dynamically change the processing capacity distribution of the preprocessing and postprocessing processors according to the rejection rate (predetermined reference value) set in STEP 5c of FIG. A simple image processing algorithm can be generated.

(Embodiment 4)
In each of the above-described embodiments, after the secondary evaluation is completed, the combination of the image processing module and the parameter value of the image processing algorithm that increases the degree of separation between the extraction target and the non-extraction target are changed, thereby changing the next evaluation target. However, the timing of creating an image processing algorithm to be evaluated is not limited to this.

  For example, as in the present embodiment, when the primary evaluation is completed, the combination of the image processing modules 42 of the image processing algorithm having a large primary evaluation value and the parameter value 43 are changed to change the next primary evaluation. A plurality of target image processing algorithms may be created. In such a case, it is possible to pass only the image processing algorithm that has been sufficiently optimized in the primary evaluation stage to the subsequent secondary evaluation processing.

  FIG. 10 shows a processing flow of the present embodiment. The difference between FIG. 10 and FIG. 9 is that the processing of STEPs 5a and 5b is provided between STEP 5 and STEP 6. Other processes are the same as those described in FIG.

  In FIG. 10, the primary evaluation routine Sj is executed for the plurality of image processing algorithms generated in STEP 1a by sharing the learning image data set 35 selected in STEP 2. In STEP 5 after the primary evaluation process, an image processing algorithm to be excluded from the evaluation target is determined and rejected based on the primary evaluation value of each routine Sj.

  In the next STEP 5a, it is determined whether or not to end the primary evaluation process. In general, in determining whether or not the primary evaluation process is completed, all of the primary evaluation values output from each routine Sj or one or more satisfy a predetermined reference value, or a predetermined time from the start of the process. If so, it is determined that the primary evaluation process is finished. Here, when the arithmetic processing capability assigned to the post-processing is sufficiently small relative to the arithmetic processing capability assigned to the pre-processing, the reference value referred to for the end evaluation of the primary evaluation in STEP 5a is increased. Alternatively, by prolonging the predetermined time, it is possible to narrow down candidates for the best image processing algorithm sufficiently at the preprocessing stage, and it is possible to generate an efficient image processing algorithm.

  If it is determined in STEP 5a that the process is finished, the secondary evaluation process in STEP 6 and subsequent steps is similarly performed. The secondary evaluation process is the same as that described with reference to FIG. On the other hand, if it is determined that the process has not ended, the process proceeds to STEP 5b.

  In STEP 5b, the combination of the image processing module 42 of the image processing algorithm having a large primary evaluation value and the parameter value 43 are changed with reference to the algorithm generation / adjustment data 44, thereby becoming the next primary evaluation target. Create multiple image processing algorithms. Then, a primary evaluation routine Sj is executed for the plurality of created image processing algorithms.

  In the present embodiment, until the primary evaluation process is determined to end the primary evaluation process in STEP 5a, a new image processing algorithm based on the previous primary evaluation result is created, and a plurality of new created evaluation processes are created. Since the current primary evaluation is executed for the image processing algorithm, and the process of executing the next primary evaluation is repeated based on the evaluation result, candidates for the best image processing algorithm are sufficiently narrowed down at the preprocessing stage. It becomes possible.

  FIG. 11 schematically shows a functional configuration for executing the processing flow of FIG. In FIG. 11, an image processing algorithm evaluation and generation apparatus includes an image data input unit 1, a RISC image processing apparatus 2A that performs post-processing, secondary evaluation, and generation of an algorithm to be evaluated, and various types of information to an operator. It comprises a display unit 3 for presentation and a DSP image processing apparatus 4A that has a plurality of DSPs and mainly performs preprocessing and primary evaluation. The DSP image processing device 4A has the number of image data conversion units 23 and the primary evaluation units 24 corresponding to the number of DSPs. Depending on the difference in processing capability between the DSP image processing apparatus 4A and the RISC image processing apparatus 2A, it is possible to set a completion determination reference referred to by the primary evaluation completion determination unit 41 that executes the processing of STEP 5a. The generation of the plurality of image processing algorithms in STEP 5b is executed by the image processing algorithm adjustment unit 42 with reference to the algorithm generation / adjustment data 44.

  Note that the environment for executing the processing flow shown in FIG. 10 is not limited to the configuration shown in FIG.

  Also in the present embodiment, in STEP 12 of FIG. 10, the primary evaluation reference (primary evaluation reference image data 12) is changed based on the secondary evaluation result of the image processing algorithm having a high evaluation value in the secondary evaluation. Therefore, even if optimization is being carried out independently for the processing related to the primary evaluation and the processing related to the secondary evaluation, the evaluation axes of the primary evaluation criteria and the secondary evaluation criteria are not different, and the degree of separation is high. It is possible to generate an image processing algorithm to achieve.

The image processing algorithm evaluation method and apparatus and the image processing algorithm generation method and apparatus of each of the embodiments described above can be applied to an image processing algorithm for product inspection if the extraction target part is a defective part of the product. The application target is not limited to defect inspection. In other words, the present invention can be applied to a case where some attention partial images are extracted from image data, or an image including a target partial image is extracted from a large number of image samples.

  In addition, the same effect can be achieved for applied image data, for example, for application to a one-dimensional signal or the like.

  In the above-described embodiment, the evaluation of the image processing algorithm is made into two stages of the primary evaluation and the secondary evaluation, the primary evaluation is performed after the conversion processing by the image processing algorithm is performed, and the feature amount is extracted after the primary evaluation is completed. The hardware configuration that implements the method of performing the secondary evaluation after carrying out the complicated processing is designed with a DSP or the like capable of high-speed filter processing up to the primary evaluation, and from the primary evaluation to the secondary evaluation. Since the RISC processor is designed with excellent versatility, the processing speed can be increased. In particular, the image processing algorithm is optimized based on the primary evaluation criteria in the pre-processing stage by the DSP, and a large number of image processing algorithms are sufficiently subjected to the primary evaluation using a plurality of DSPs. By applying only the algorithm to the next step performed by the RISC processor, a sufficiently high-speed image processing algorithm can be generated. Here, even if optimization is advanced independently by the DSP by changing the standard of the primary evaluation based on the primary evaluation result of the image processing algorithm having a high evaluation result in the secondary evaluation. There is no difference between the evaluation axis of the standard and the standard of the secondary evaluation, and it is possible to generate an image processing algorithm that achieves a high degree of separation.

(Embodiment 5)
In each of the above-described embodiments, the original image data 36 to be processed is stored in advance in the memory 324 as the learning image data set 35. However, each time processing is executed, it is input from the outside. Also good. That is, in the configuration of FIG. 3, various types of information can be transmitted / received to / from the outside via the communication network 109, so that the received image may include the original image data 36 or the reference image data 37 to be processed. The original image data 36 may be data of an image taken by an external camera, for example.

  As a recording medium on which data such as a program code for realizing the functions described in the above embodiments is recorded, a memory required for processing by the computer shown in FIG. The recording medium itself may be a program medium, or a magnetic tape device (not shown) and a program reading device such as a CD-ROM driving device 340 are provided as an external storage device, and a magnetic tape (not shown) as a recording medium is provided there. It may be a program medium that can be read by inserting the CD-ROM 342. In any case, the stored program may be configured to be accessed and executed by the CPU 312. In any case, the program is once read from the recording medium, and the read program is shown in FIG. The program may be loaded into a predetermined program storage area of the computer, for example, the program storage area of the memory 324, and read and executed by the CPU 312. This loading program is assumed to be stored in advance in the computer.

Here, the above-described program medium is a recording medium configured to be separable from the computer main body, and may be a medium that carries the program fixedly. For example, tape systems such as magnetic tape and cassette tape, magnetic disks such as FD332 and fixed disk 326, CD-ROM342 / MO (Magnetic Optical Disc) / MD (Mini Disc) / DVD
(Digital Versatile Disc) and other optical discs, IC cards (including memory cards) / optical cards, etc., or mask ROM, EPROM (Erasable and Programmable ROM), EEPROM (Electrically EPROM), flash ROM, etc. A semiconductor memory would be applicable.

In the present embodiment, since the computer is connected to the communication network 109 including the Internet via the communication interface 380, the program may be downloaded from the communication network 109. When the program is downloaded from the communication network 109 in this way, the download program is stored in the computer main body in advance, or is installed in the computer main body from another recording medium in advance.

  Note that the content stored in the recording medium is not limited to a program, and may be data.

  Further, when the computer executes the program code read from these recording media, not only the functions shown in FIGS. 1 and 2 are realized, but also an OS running on the computer based on the instruction of the program code. It goes without saying that (operating system) or the like performs part or all of the actual processing, and these functions are realized by the processing.

  Further, after the program code read from the recording medium is written in a memory provided in a function expansion board inserted in the computer or a function expansion unit connected to the computer, the function expansion is performed based on the instruction of the program code. It goes without saying that the CPU or the like provided in the board or function expansion unit performs part or all of the actual processing and the above-described functions are realized by the processing.

  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.

3 is a process flowchart according to the first embodiment. 2 is a functional configuration diagram of an apparatus according to Embodiment 1. FIG. It is a block diagram of the computer by which the apparatus which concerns on each embodiment is mounted. (A)-(C) is a figure explaining the change of the combination of the image processing module which concerns on each embodiment. 6 is a diagram for explaining application of an image processing algorithm evaluation method to a learning image data set according to Embodiment 1. FIG. 10 is a process flowchart according to the second embodiment. FIG. 7 is a diagram illustrating a case where the procedure of FIG. 6 is applied to an actual learning image data set. FIG. 6 is a functional configuration diagram of an apparatus according to a third embodiment. 10 is a process flowchart according to the third embodiment. 10 is a process flowchart according to the fourth embodiment. FIG. 6 is a functional configuration diagram of an apparatus according to a fourth embodiment.

Explanation of symbols

12 primary evaluation reference data, 13 post-conversion image data, 14 image processing algorithm data, 21 extraction target information setting unit, 22 image processing algorithm generation / adjustment unit, 23 image data conversion unit, 24 primary evaluation unit, 25 rejection determination 26, converted image data storage processing unit, 27 feature quantity calculation unit, 28 classification unit, 29 separability calculation unit, 30 secondary evaluation unit, 31 image processing algorithm storage processing unit, 32 primary evaluation criterion change unit, 33 Image processing algorithm generation completion determination unit, 34 learning image data storage unit, 35 learning image data set, 36 original image data, 37 reference image data, 44 algorithm generation / adjustment data.

Claims (34)

The image data of the processing target image is converted so that it can be processed, and one or more true defect portions are processed using the feature amounts calculated for a plurality of defect candidate portions in the image based on the converted image data. An image processing algorithm to be extracted from an image to be evaluated using image data having predetermined information regarding the one or more true defect portions to be extracted included in the processing target image,
A conversion step of performing conversion processing of the image processing algorithm on the processing target image data;
A primary evaluation step of evaluating the ability of the conversion processing based on the post-conversion image data converted so as to be processed obtained by applying the conversion step and the predetermined information;
A feature amount calculating step of calculating one or more types of feature amounts for each of the plurality of defect candidate portions of the processing target image;
A classifying step of classifying the plurality of defect candidate portions whose feature values are calculated by the feature value calculating step based on the predetermined information into the true defect portions and the rest as pseudo defect portions;
Calculate the degree of separation in the feature space defined by the one or more feature quantities of the one or more feature quantity groups of the true defect section and the one or more feature quantity groups of the pseudo defect section. A degree-of-separation calculation step,
An image processing algorithm evaluation method comprising: a secondary evaluation step of evaluating performance of the image processing algorithm based on at least one of the degree of separation calculated in the separation degree calculation step and an evaluation result of the primary evaluation step.   The image processing algorithm evaluation method according to claim 1, further comprising a rejection step of rejecting an image processing algorithm that does not satisfy the evaluation criteria based on an evaluation result in the primary evaluation step.   The image processing algorithm evaluation method according to claim 1, further comprising a primary evaluation standard changing step of changing the evaluation standard of the primary evaluation step based on an evaluation result of the secondary evaluation step. The image data having the predetermined information is
4. The apparatus according to claim 1, comprising primary evaluation reference image data expressing the one or more true defect portions to be extracted included in the processing target image and other pseudo defect portions with different density values. The image processing algorithm evaluation method according to the item. In the primary evaluation process,
The image processing algorithm evaluation method according to claim 4, wherein evaluation is performed by comparing the primary evaluation reference image data and the converted image data. In the primary evaluation standard changing step,
The image processing algorithm evaluation method according to claim 4 or 5, wherein the primary evaluation reference image data is changed using the converted image data converted by the image processing algorithm that has been evaluated in the secondary evaluation step. . In the primary evaluation standard changing step,
The image processing algorithm evaluation method according to claim 6, wherein the converted image data converted by the image processing algorithm that obtains the best evaluation in the secondary evaluation step is replaced with the primary evaluation reference image data. The image data of the processing target image is converted so that it can be processed, and one or more true defect portions are processed using the feature amounts calculated for a plurality of defect candidate portions in the image based on the converted image data. A method for generating the image processing algorithm based on an evaluation using image data having predetermined information regarding the one or more true defect portions to be extracted included in an image, for an image processing algorithm extracted from an image,
A conversion step of performing conversion processing of the image processing algorithm to be input on the image data to be processed;
A primary evaluation step for evaluating the capability of the conversion processing based on the post-conversion image data converted so as to be processed obtained by applying the conversion step and the predetermined information;
A feature amount calculating step of calculating one or more types of feature amounts for each of the plurality of defect candidate portions of the processing target image;
A classifying step of classifying the plurality of defect candidate portions whose feature values are calculated by the feature value calculating step based on the predetermined information into the true defect portions and the rest as pseudo defect portions;
Calculate the degree of separation in the feature space defined by the one or more feature quantities of the one or more feature quantity groups of the true defect section and the one or more feature quantity groups of the pseudo defect section. A degree-of-separation calculation step,
A secondary evaluation step of evaluating the performance of the image processing algorithm based on at least one of the separation degree calculated by the separation degree calculation step and an evaluation result of the primary evaluation step;
A completion determination step for determining whether the generation of the image processing algorithm is completed based on a predetermined criterion;
When it is determined that the completion is not completed by the completion determination step, by changing an element of the image processing algorithm based on the evaluation result output by at least one of the primary evaluation step and the secondary evaluation step. An image processing algorithm generation method comprising: an adjustment step of creating one or more image processing algorithms and outputting the created image processing algorithms to the conversion step.   The image processing algorithm generation method according to claim 8, further comprising a rejection step of rejecting an image processing algorithm that does not satisfy an evaluation criterion based on an evaluation result in the primary evaluation step. In the rejection step,
The image processing algorithm according to claim 9, wherein the image processing algorithm alone is rejected according to a rejection rate determined based on the processing capability after the rejection step, so that the processing in the subsequent steps only for the image processing algorithm is not performed. Generation method. In the rejection step,
By rejecting the image processing algorithm in accordance with the rejection rate determined based on the correlation state between the evaluation result of the primary evaluation step and the evaluation result of the secondary evaluation step, only the image processing algorithm in the subsequent steps The image processing algorithm generation method according to claim 9, wherein the processing is not performed.   The image processing algorithm generation method according to any one of claims 8 to 11, further comprising a primary evaluation standard changing step of changing an evaluation standard of the primary evaluation step based on an evaluation result of the secondary evaluation step. . The image data having the predetermined information is
The primary evaluation reference image data expressing the one or more true defect portions to be extracted included in the processing target image and other pseudo defect portions with different density values, respectively. The image processing algorithm generation method according to Item. In the primary evaluation process,
The image processing algorithm generation method according to claim 13, wherein the evaluation is performed by comparing the primary evaluation reference image data and the converted image data. In the primary evaluation standard changing step,
The image processing according to claim 13 or 14, wherein the primary evaluation reference image data is changed using the converted image data converted by the image processing algorithm that has been evaluated in the secondary evaluation step. Algorithm generation method. In the primary evaluation standard changing step,
The image processing algorithm generation method according to claim 15, wherein the converted image data converted by the image processing algorithm that obtains the best evaluation in the secondary evaluation step is replaced with the primary evaluation reference image data.   A program for causing a computer to execute the image processing algorithm evaluation method according to any one of claims 1 to 7.   A computer-readable recording medium on which a program for causing a computer to execute the image processing algorithm evaluation method according to any one of claims 1 to 7 is recorded.   A program for causing a computer to execute the image processing algorithm generation method according to any one of claims 8 to 16.   The computer-readable recording medium which recorded the program for performing the image processing algorithm production | generation method of any one of Claim 8 to 16 by computer. The image data of the processing target image is converted so that it can be processed, and one or more true defect portions are processed using the feature amounts calculated for a plurality of defect candidate portions in the image based on the converted image data. An image processing algorithm to be extracted from an image to be generated based on an evaluation using image data having predetermined information regarding the one or more true defect portions to be extracted included in the image. And
Conversion means for performing conversion processing of the image processing algorithm to be input on image data of the processing target image;
Primary evaluation means for evaluating the ability of the conversion processing based on the post-conversion image data converted so as to be processed obtained by applying the conversion means and the predetermined information;
A feature amount calculating means for calculating one or more types of feature amounts for each of the plurality of defect candidate portions of the processing target image;
Classifying means for classifying the plurality of defect candidate parts, whose feature values are calculated by the feature value calculating means based on the predetermined information, into the true defect part and the rest as pseudo defect parts;
Calculate the degree of separation in the feature space defined by the one or more feature quantities of the one or more feature quantity groups of the true defect section and the one or more feature quantity groups of the pseudo defect section. A degree of separation calculating means,
Secondary evaluation means for evaluating the performance of the image processing algorithm based on at least one of the separation degree calculated by the separation degree calculation means and the evaluation result of the primary evaluation step;
Completion determination means for determining whether generation of the image processing algorithm is completed based on a predetermined criterion;
If it is determined by the completion determining means that the image has not been completed, 1 is obtained by changing the elements of the image processing algorithm based on the evaluation result output by at least one of the primary evaluation means and the secondary evaluation means. An image processing algorithm generation apparatus comprising: an adjustment unit that generates one or more image processing algorithms and outputs the generated image processing algorithms to the conversion unit. The image processing algorithm generation device according to claim 21, further comprising a rejection unit that rejects an image processing algorithm that does not satisfy an evaluation criterion based on an evaluation result of the primary evaluation unit. In the rejection means,
23. The image processing algorithm according to claim 22, wherein the image processing algorithm is rejected according to a rejection rate determined on the basis of the processing capability after the rejection means, so that the processing by the means subsequent to the image processing algorithm alone is not performed. Generator. In the rejection means,
By rejecting the image processing algorithm in accordance with the rejection rate determined based on the correlation state between the evaluation result by the rejection means and the evaluation result by the secondary evaluation means, only the image processing algorithm is processed by the subsequent means. The image processing algorithm generation device according to claim 22, which is not implemented.   The image processing algorithm generation device according to any one of claims 21 to 24, further comprising a primary evaluation standard changing unit that changes an evaluation standard of the primary evaluation unit based on an evaluation result of the secondary evaluation unit. .   The two or more conversion means and a series of means including the feature quantity calculation means, the classification means, and the separation degree calculation means are provided, and evaluations of two or more image processing algorithms are processed in parallel. 26. The image processing algorithm generation device according to any one of 25. Primary completion determination means for determining whether primary evaluation is completed based on the evaluation result output by the primary evaluation means;
One or more images by changing any element of the image processing algorithm based on the evaluation result output by the primary evaluation unit when the primary evaluation completion determination unit determines that the image has not been completed. 27. The image processing algorithm generation apparatus according to claim 26, further comprising: a primary adjustment unit that generates a processing algorithm and outputs the generated image processing algorithm to the conversion unit. The image data of the processing target image is converted so that it can be processed, and one or more true defect portions are processed using the feature amounts calculated for a plurality of defect candidate portions in the image based on the converted image data. An image processing algorithm to be extracted from an image to be evaluated using image data having predetermined information regarding the one or more true defect portions to be extracted included in the processing target image,
Conversion means for performing conversion processing of the image processing algorithm on the processing target image data;
Primary evaluation means for evaluating the ability of the conversion processing based on the post-conversion image data converted so as to be processed obtained by applying the conversion means and the predetermined information;
A feature amount calculating means for calculating one or more types of feature amounts for each of the plurality of defect candidate portions of the processing target image;
Classifying means for classifying the plurality of defect candidate parts, whose feature values are calculated by the feature value calculating means based on the predetermined information, into the true defect part and the rest as pseudo defect parts;
Calculate the degree of separation in the feature space defined by the one or more feature quantities of the one or more feature quantity groups of the true defect section and the one or more feature quantity groups of the pseudo defect section. A degree of separation calculating means,
An image processing algorithm evaluation apparatus comprising: secondary evaluation means for evaluating the performance of the image processing algorithm based on at least one of the degree of separation calculated by the degree of separation calculation means and the evaluation result of the primary evaluation means. 29. The image processing algorithm evaluation apparatus according to claim 28, further comprising: a rejection unit that rejects an image processing algorithm that does not satisfy an evaluation criterion based on an evaluation result of the primary evaluation unit.   30. The image processing algorithm evaluation apparatus according to claim 28, further comprising primary evaluation standard changing means for changing the evaluation standard of the primary evaluation means based on an evaluation result of the secondary evaluation means. The image data having the predetermined information is
31. The apparatus according to any one of claims 28 to 30, including primary evaluation reference image data in which the one or more true defect portions to be extracted and other pseudo defect portions to be extracted included in the processing target image are expressed by different density values. The image processing algorithm evaluation apparatus according to Item. In the primary evaluation means,
32. The image processing algorithm evaluation apparatus according to claim 31, wherein evaluation is performed by comparing the primary evaluation reference image data and the converted image data. In the primary evaluation standard changing means,
The image processing algorithm evaluation apparatus according to claim 31 or 32, wherein the primary evaluation reference image data is changed using the converted image data converted by the image processing algorithm that has been evaluated in the secondary evaluation unit. . In the primary evaluation standard changing means,
34. The image processing algorithm evaluation device according to claim 33, wherein the converted image data converted by the image processing algorithm that obtains the best evaluation in the secondary evaluation means is replaced with the primary evaluation reference image data.

Images