JP2021131835A - Image processing system and image processing program - Google Patents

Abstract

To enable reduction of a learning time while maintaining performance of image processing algorithm in automatic generation of the image processing algorithm using evolutionary calculation.SOLUTION: An image processing system includes: an individual group storage unit that stores a plurality of pieces of individual information defining execution order of a plurality of function units; an image storage unit that stores a plurality of learning images; an image processing unit that processes the learning image on the basis of the individual information to generate corresponding output information; a teacher information storage unit that stores teacher information for evaluating the output information; an evaluation value calculation unit that compares the output information with the teacher information to calculate an evaluation value for each individual information and each learning image to rank the individual information; a gene operation unit that selects or changes the individual information on the basis of ranking information of the individual information to update the individual group storage unit; and a learning image reduction unit that reduces the learning image on the basis of the evaluation value for each learning image.SELECTED DRAWING: Figure 3

Description

The present invention relates to an image processing technique such as finding a detection target from an image, and particularly to an image processing system for optimizing image processing.

Conventionally, in the automatic generation of an image processing algorithm such as finding a detection target from an image, image processing is performed on a learning image using individual information that defines an execution order of a plurality of functional units, and the obtained output image is used. The evaluation value is calculated by comparing with the teacher image, and the result is reflected in the individual information for learning to optimize the image processing.
In the automatic generation of such an image processing algorithm, there is a problem that the number of images used for learning increases and the learning time becomes long depending on the type of detection target.

Specifically, when the detection target has various variations such as scratches, a learning image covering these variations is required. For this reason, in the automatic generation of an image processing algorithm for finding a detection target having such various variations, learning is performed using a large number of images, and there is a problem that it takes a long time to optimize the image processing algorithm.
On the other hand, if the number of learning images to be detected, which requires coverage of variations, is simply reduced, there arises a problem that the performance of the generated image processing algorithm deteriorates. Therefore, it has not been easy to reduce the number of learning images while maintaining the performance of the image processing algorithm.

Japanese Patent No. 6102947

Here, as an image processing technique related to the reduction of learning images, the image processing apparatus described in Patent Document 1 can be mentioned.
It is described that in this apparatus, clustering is performed based on the feature amount of an image at the time of initial learning, and a representative image is selected from each cluster to reduce the number of images.

However, in this technique, since clustering is performed at the time of initial learning, the selected image is fixed. In addition, it is complicated because it is necessary to calculate the feature amount of the image in advance for the purpose of selecting the image. Further, when the feature amount capable of accurately performing clustering is known by this method, the detection target of the image processing is a simple one with less need to perform machine learning.
For this reason, an image processing system that can more effectively reduce learning images while maintaining the performance of the image processing algorithm when automatically generating an image processing algorithm that finds a detection target that requires coverage of variations. Was required.

Therefore, the present inventors have conducted diligent research, and in the automatic generation of an image processing algorithm using evolutionary calculation, the output image obtained by performing image processing on the training image was calculated by comparing it with the teacher image. The present invention has been completed by making it possible to select an image to be used in learning based on the evaluation value and succeeding in shortening the learning time while maintaining the performance of the generated image processing algorithm.

That is, according to the present invention, since the evaluation value calculated at the time of learning is used, it is not necessary to set the feature amount at the initial stage. Further, since the evaluation value for the image changes as the learning progresses, the image used for the learning is not fixed. Furthermore, since the evaluation value represents the characteristics of the image processing algorithm generated by evolutionary computation, it is possible to expect an improvement in the performance of the generated image processing algorithm.

The present invention has been made in view of the above circumstances, and in automatic generation of an image processing algorithm using evolutionary calculation, image processing capable of shortening the learning time while maintaining the performance of the image processing algorithm. The purpose is to provide a system and an image processing program.

In order to achieve the above object, the image processing system of the present invention is an image processing system that optimizes image processing by using genetic manipulation, which is an optimization method that mathematically simulates the evolution of living organisms, and is a linear image processing system. Based on the individual information, a population storage unit that stores a plurality of individual information that defines an execution order of a plurality of functional units in a structure, a tree structure, or a network, an image storage unit that stores a plurality of learning images, and the individual information. An image processing unit that processes the learning image and generates output information corresponding to the learning image, and a teacher information storage unit that stores teacher information corresponding to the learning image for evaluating the output information. , The output information and the teacher information are compared, an evaluation value is calculated for each individual information and each learning image, and the evaluation value calculation for ranking the individual information based on the evaluation value for each individual information. The image, the gene manipulation unit that updates the population storage unit by selecting or changing the individual information based on the ranking information of the individual information, and the evaluation value for each learning image. It is configured to include a learning image reduction unit that reduces the learning image processed by the processing unit.

Further, in the image processing system of the present invention, the gene manipulation unit updates the population storage unit to create a plurality of individual information of a new generation, and the image processing unit uses the gene manipulation unit. When the individual information of a new generation is created, the individual information of the new generation is used as the individual information, and the learning image reduction unit performs the learning after the set generation or for each set generation. Based on the evaluation value for each image, a part of the images is selected from the learning images stored in the image storage unit, and the image processing unit uses the selected part of the images for learning. It is preferable that the configuration is used as an image.

Further, in the image processing system of the present invention, the learning image reduction unit clusters the learning images after the set generation or for each set generation based on the evaluation value of the learning image. , It is also preferable to select a part of the learning images from each class.

It is also preferable that the image processing system of the present invention has a configuration in which the learning image reduction unit clusters the learning images by using a discriminant analysis method, a hierarchical cluster analysis, or a non-hierarchical cluster analysis.

Further, in the image processing system of the present invention, the learning image reduction unit uses the learning image reduction unit after the set generation or for each set generation based on the rank or correlation coefficient of the evaluation value of the learning image. It is also preferable that the learning image of the part is selected.

Further, the image processing program of the present invention is an image processing system that optimizes image processing by using genetic manipulation, which is an optimization method that mathematically simulates the evolution of living organisms. A population storage unit that stores a plurality of individual information that defines an execution order of a plurality of functional units in a tree structure or a network, an image storage unit that stores a plurality of learning images, and the learning unit based on the individual information. An image processing unit that processes an image and generates output information corresponding to the learning image, a teacher information storage unit that stores teacher information corresponding to the learning image for evaluating the output information, and the output information. An evaluation value calculation unit that compares the teacher information, calculates an evaluation value for each individual information and each learning image, and ranks the individual information based on the evaluation value for each individual information. It is processed by the gene manipulation unit that updates the population storage unit by selecting or changing the individual information based on the ranking information, and the image processing unit based on the evaluation value for each learning image. It is configured to function as a learning image reduction unit that reduces learning images.

According to the present invention, in the automatic generation of an image processing algorithm using evolutionary calculation, it is possible to provide an image processing system and an image processing program capable of shortening the learning time while maintaining the performance of the image processing algorithm. It will be possible.

It is a block diagram which shows the structure of the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of selection using the constant by the image reduction part for learning in the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of selection using the linear function by the image reduction part for learning in the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of selection using the sigmoid function by the image reduction part for learning in the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of selection using the rule by the image reduction part for learning in the image processing apparatus of 1st Embodiment of this invention. It is a flowchart which shows the processing procedure by the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of the evaluation value data in the evaluation result storage part in the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of the ranking data in the evaluation result storage part in the image processing apparatus of 1st Embodiment of this invention. It is a flowchart which shows the processing procedure of a gene manipulation in the processing procedure by the image processing apparatus of 1st Embodiment of this invention. It is explanatory drawing of the evaluation value data of a plurality of generations in the evaluation result storage part in the image processing apparatus of the 1st Embodiment of this invention. It is a block diagram which shows the structure of the image processing apparatus of the 2nd Embodiment of this invention. It is a flowchart which shows the processing procedure by the image processing apparatus of the 2nd Embodiment of this invention. It is a figure which shows the background image in the learning image and the verification image used in an Example and a reference example. It is a figure which shows the copy image of arbitrary position of the background image in the learning image and the verification image used in an Example and a reference example. It is a figure which shows the sample image 1 (the image which both the detection target and the non-detection target are not drawn) in the learning image and the verification image used in an Example and a reference example. It is a figure which shows the sample image 2 (the image which the detection target was drawn) in the learning image and the verification image used in an Example and a reference example. It is a figure which shows the sample image 3 (the image which the non-detection target was drawn) in the learning image and the verification image used in an Example and a reference example. It is a figure which shows various parameters used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: noise removal) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: luminance correction) of a node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: binarization) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: edge detection) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: frequency filter) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: calculation) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: morphology) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: detection target determination) of the node used in an Example and a reference example. It is explanatory drawing which shows the function (image processing: branch) of a node used in an Example and a reference example. Represents the learning time, learning result (evaluation value), and verification result (evaluation value) according to Examples (reduction rate 20%, 60%, 80%, 90%, 95%) and Reference Example 2 (reduction rate 0%). It is a figure which shows the graph. The learning time, learning result (evaluation value), and verification result (evaluation value) according to Reference Example 1 (reduction rate 20%, 60%, 80%, 90%, 95%) and Reference Example 2 (reduction rate 0%) It is a figure which shows the graph which represents. It is a figure which shows the graph which compared the learning result (evaluation value) and the verification result (evaluation value) by an Example and a reference example.

Hereinafter, an image processing system of the present invention and an embodiment of an image processing program will be described in detail. However, the present invention is not limited to the specific contents of the following embodiments and examples described later.

[First Embodiment]
First, the image processing system and the image processing program according to the first embodiment of the present invention will be described with reference to FIGS. 1 and 2A to 2D. FIG. 1 is a block diagram showing a configuration of an image processing device corresponding to the image processing system of the present embodiment. 2A to 2D are explanatory views of selection using a constant, selection using a linear function, selection using a sigmoid function, and selection using a rule by the learning image reduction unit in the image processing apparatus, respectively.

The image processing system of the present embodiment optimizes image processing by using genetic manipulation, which is an optimization method that mathematically simulates the evolution of living organisms. Examples of this image processing include an image processing algorithm for detecting a detection target from an image, an image noise removal algorithm, an image high resolution algorithm, an image composition algorithm, and the like.

Specifically, as shown in FIG. 1, the image processing device 1 of the present embodiment has a functional unit storage unit 10, an individual generation unit 11, a population storage unit 12, an image input unit 13, an image storage unit 14, and a teacher. It includes an information input unit 15, a teacher information storage unit 16, an image processing unit 17, an output information storage unit 18, an evaluation value calculation unit 19, an evaluation result storage unit 20, a gene manipulation unit 21, and an image reduction unit 22 for learning. ..
As shown in FIG. 1, all of these configurations in the image processing device 1 of the present embodiment can be provided in one information processing device. Further, each of these configurations may be distributed and provided in each device of the image processing system including a plurality of information processing devices. This also applies to the second embodiment described later.

The functional unit storage unit 10 stores nodes (corresponding to genes), which are functional units that execute a certain process, for each node number. In this embodiment, various nodes for performing image processing can be used as this node, and in the examples described later, those shown in FIGS. 13A to 13I are used.
Further, when a parameter used for executing each process exists, these nodes are stored in the functional unit storage unit 10 together with the parameter.

Further, as the node, a node that does not execute any processing may be used, or a node (node set) that is a collection of functional units that execute a plurality of processes may be used. By using the nodes including those that do not perform any processing, the actual number of nodes can be reduced, for example, when the set value of the initial number of nodes for generating individual information is too large.

In the present specification, "image processing" includes various filter processing for noise removal, correction processing such as luminance correction, binarization processing, edge detection processing, frequency filter processing, arithmetic processing such as four-rule calculation, and morphology. Processing, detection target determination processing, branching processing, etc. are included.

The individual generation unit 11 generates individual information composed of an array of a plurality of nodes by using the nodes stored in the functional unit storage unit 10. At this time, the individual generation unit 11 can randomly generate a plurality of individual information based on the plurality of nodes. This individual information constitutes an image processing algorithm.
At this time, the individual generation unit 11 can define the execution order of a plurality of nodes in a linear structure, a tree structure, or a network in the individual information. The network is a directed network composed of oriented links and can include feedback.

It should be noted that if network-like individual information is used, learning can be performed including a network structure having more flexible expressive power as compared with the case of using individual information consisting only of a linear structure or a tree structure, which is optimal. It has an advantage in deriving solutions (optimal practical solutions, practical solutions). That is, it is difficult for a person to know in advance which structure is the optimum solution for a problem, so it is difficult to determine the conditions of the structure before machine learning. Therefore, searching for the optimal solution under a wide range of conditions using the individual information that defines the execution order of multiple nodes in a network that can include a cycle may lead to the true optimal solution. It is considered to be effective in enhancing.

The population storage unit 12 stores a population composed of a plurality of individual information generated by the individual generation unit 11. As this population, one generation or a plurality of generations can be stored in the population storage unit 12, and individual information can be stored in the population storage unit 12 for each generation number.
The evaluation value calculation unit 19, which will be described later, can rank individual information for each generation or for each set generation, and can store the ranking information in the evaluation result storage unit 20.

The image input unit 13 inputs a plurality of learning images and a plurality of verification images to the image processing device 1, and stores the identification information of the images in the image storage unit 14.
The image for learning is processed by all individual information (image processing algorithm) to obtain output information, and the evaluation value is calculated by comparing the output information with the teacher information, and the evaluation value is used for the individual. It can be used to carry out ranking.
The verification image is processed by the best individual information among the predetermined individual information to obtain the output information, and the evaluation value is calculated by comparing the output information with the teacher information, and the evaluation value is used. It can be used to verify the validity of the analysis model and whether over-learning has occurred.

Since the verification image is used by a person to analyze the result, it basically does not affect the evolutionary computation, but when the evaluation value of the verification image is used for the end judgment, for example, the verification image. Automatically affects evolutionary computation if it ends when the evaluation value of is reached the threshold, or if it ends early when there is a large difference between the evaluation values of the learning image and the verification image. become.

The teacher information input unit 15 inputs teacher information for evaluating the output information, and stores the identification information of the corresponding image in the teacher information storage unit 16.
As the teacher information, a teacher image corresponding to a plurality of learning images, a teacher image corresponding to a plurality of verification images, and the like can be used.
In the image processing system of the present embodiment, the teacher information is not limited to the image, and the number and position of the detection targets in the image, the range in which the detection target exists, and the like can be used as the teacher information.

The image processing unit 17 inputs an image from the image storage unit 14, and also inputs individual information from the population storage unit 12, and sequentially executes a plurality of nodes included in the individual information based on the individual information. The image processing unit 17 also executes a node for executing each of the image input processing and the output image storage processing.

At this time, each node executes the process corresponding to each function by using the information obtained from the image. Further, each node can selectively execute the next node corresponding to the processing result based on the processing result. Therefore, not all nodes are necessarily executed in image processing even if the nodes are included in the individual information. Of course, all the nodes in the individual information may be executed. Further, by giving feedback, a node defined as only one in one individual information may be executed a plurality of times.

In addition, as described above, the node functions include various filter processing for noise removal, correction processing such as luminance correction, binarization processing, edge detection processing, frequency filter processing, and arithmetic processing such as four-rule calculation. There is image processing in a narrow sense such as morphology processing and branch processing.

Further, in the detection target determination process as a function of the node, the image processing unit 17 labels the detection target candidates in a certain range in the image, calculates the feature amount, and provides the feature amount based on the threshold value or the like. It is performed as a process of determining whether or not the detection target candidate is a correct detection target.
In the detection target determination process, it is also possible to perform a process of labeling a non-detection target candidate, calculating the feature amount thereof, and determining that the non-detection target candidate is not a detection target because the feature amount is not provided based on a threshold value or the like. ..

Further, the image processing unit 17 creates an image as output information for each of the learning image and the verification image, and outputs information storage unit 18 for each individual information (identification information) and for each image (identification information). Can be memorized in. When the teacher information is the number, position, or range in which the detection target exists in the image, the output information includes the number, position, or range in which the detection target exists, respectively. It is created and stored in the output information storage unit 18.

When the image processing unit 17 creates an image as output information for the purpose of finding a detection target from the image, if a detection target is detected for each of the learning image and the verification image, the detection target exists. An image to which information indicating the area to be detected is added is created, and if the detection target is not detected, an image to which the information is not added is created.
Then, this output information can be stored in the output information storage unit 18 for each individual information and each image.

When a detection target exists in the learning image and the verification image, the evaluation value calculation unit 19 can be made to perform a function of generating an image or the like to which information indicating the region is added. That is, after image processing is performed by the image processing unit 17, the evaluation value calculation unit 19 generates an image to which information indicating the area to be detected is added, stores the image in the output information storage unit 18, and uses this to store the evaluation value. It is also possible to have a configuration that calculates.

The evaluation value calculation unit 19 inputs output information from the output information storage unit 18, and also inputs teacher information corresponding to the output information from the teacher information storage unit 16.
Then, the evaluation value is calculated by comparing the output information and the teacher information, and the obtained evaluation value can be stored in the evaluation result storage unit 20 for each individual information and each image.

The evaluation value calculation unit 19 can calculate the evaluation value by the following formula, for example, using the mean square error.

Further, the evaluation value calculation unit 19 finishes calculating the evaluation value for all the individual information (or the individual information for one generation) stored in the population storage unit 12, and stores the evaluation value in the evaluation result storage unit 20. For example, individual information can be ranked (ranked) based on those evaluation values, and the obtained ranking information can be stored in the evaluation result storage unit 20.
Further, the evaluation value calculation unit 19 can rank (rank) the learning images for each individual information based on the evaluation value, and store the obtained ranking information in the evaluation result storage unit 20.

Further, the evaluation value calculation unit 19 can perform an end determination for determining whether or not to end the image processing by the image processing unit 17. The end determination can be performed based on whether or not the end condition is satisfied, and if the end condition is satisfied, the image processing by the image processing unit 17 can be terminated and the processing by the image processing device 1 can be terminated. Then, the individual information having the highest evaluation value stored in the evaluation result storage unit 20 is obtained as the optimized individual information for performing image processing.

Examples of the termination condition include the case where image processing is completed for a preset population of all generations, and the case where evolution has not occurred in a certain number of generations or more.
The case where evolution has not occurred is the case where the evaluation value is not improved from the time of comparison.
The end determination function may be separated from the evaluation value calculation unit 19, and the image processing device 1 of the present embodiment may include an end determination unit for performing the end determination.

The gene manipulation unit 21 updates the population storage unit 12 by selecting or changing the individual information based on the ranking information of the individual information in the evaluation result storage unit 20.
As a result, individual information of a new generation population can be sequentially added to the population storage unit 12.

Specifically, the gene manipulation unit 21 includes an elite individual storage unit that leaves individual information with a high evaluation value to the next generation, and an individual selection unit that leaves individual information to the next generation with a certain probability based on the evaluation value. It can have a crossing part that exchanges a part of two individual information with each other, and a mutation part that randomly rewrites a part or all of one selected individual information or a node parameter.

The elite individual preservation unit can unconditionally leave the individual information with the best evaluation value in the population of each generation to the next generation.
The individual selection unit can select individual information having a high evaluation value in the population of each generation with a high probability and perform processing such as leaving it to the next generation.

As a crossover process, the crossover portion can perform one-way crossover in which the same (node number) arrangement in the array of each node of a plurality of individual information is replaced with a certain probability. In addition, as a crossover process, the crossover part is a one-point crossover in which one point in the same sequence is used as a boundary in the array of each node and one of the sequences is exchanged with each other, or an array is mutually used as a boundary in a plurality of points in the same sequence. It is also possible to perform multi-point crossover to be replaced. The crossing section can perform such crossing processing on the individual information selected by the individual selection section.

As a mutation process, the mutation unit rewrites a part of the individual information or rewrites all of the individual information with a certain probability for each individual information using the node stored in the functional unit storage unit 10. be able to. The mutation unit can perform such mutation processing on the individual information selected by the individual selection unit. In addition, the mutation unit can also generate new individual information by using the node stored in the functional unit storage unit 10 as the mutation process. Further, it is also possible to cause the mutation portion to perform parameter mutation as a mutation process by randomly rewriting a part or all of the parameters of the node of the selected parent individual information.

If the number of homologous individuals generated based on the same individual information increases, diversity will be lost and evolution will enter a dead end (a state in which a mathematically local solution has fallen). On the other hand, if the number of individuals other than the same series is increased too much, it approaches random search and efficient evolution is not performed.
Therefore, in order to have diversity in the early stage of learning, processing by the mutation part is performed so as to increase the number of individuals other than the same series, and in the middle stage of learning, the same series is processed by the individual selection part in order to carry out efficient evolution. It is preferable to increase it. In addition, when evolution is stopped, it is possible to preferably increase the number of mutations other than the same series by the mutation site in order to give diversity again.

Then, the gene manipulation unit 21 generates a new generation population and stores it in the population storage unit 12. At this time, the genetic manipulation unit 21 can update the population storage unit 12 by overwriting the population of the previous generation. In addition, the population storage unit 12 can be updated by storing a population of a plurality of generations for each generation.

The learning image reduction unit 22 determines whether or not to execute the learning image reduction process. This determination criterion can be set as appropriate, a predetermined set value can be set, and the learning image reduction process can be executed for all generations (after the set generation) after the set value. In addition, the reduction process can be executed for each set number of generations.

Further, when the learning image reduction unit 22 determines that the reduction process is to be executed, the learning is image-processed by the image processing unit 17 based on the evaluation value of each learning image stored in the evaluation result storage unit 20. Reduce images for use.
The learning image reduction unit 22 can select a learning image by various methods based on the evaluation value of each learning image in the learning image reduction process.
Specifically, as a method for reducing learning images, "1. Reduction by ranking of evaluation values", "2. Reduction using correlation coefficient of evaluation values", and "3. Reduction by clustering" can be mentioned. can.

1. 1. Reduction by ranking of evaluation values (1) Reduction by constant Selection is made with respect to the ranking of evaluation values using the selection probability expressed by the following equation and the constant shown in FIG. 2A. In this case, it can be reduced evenly with respect to the ranking.

(2) Reduction by linear function As shown in the following equation and FIG. 2B, the evaluation value rank is weighted and selected. In this case, when the slope of the straight line is +, many bad images are reduced, and when-, many images with good evaluation values are reduced. It is also possible to set the order not to be selected by the constant b.

(3) Reduction using the sigmoid function As shown in the following equation and FIG. 2C, the evaluation value rank is reduced by using the weight by the sigmoid function. Similar to the selection by the linear function, an image having a good evaluation value or an image having a bad evaluation value can be significantly selected.

(3) Others In addition, exponential functions, logarithmic functions, etc. can also be used. Further, as shown in FIG. 2D, it is also possible to arbitrarily select a certain rank or higher or lower, and between them by using a predetermined rule.

2. Reduction using the correlation coefficient of the evaluation value The number of images used for learning is reduced by calculating the correlation coefficient with respect to the evaluation value of the image and reducing one of the images having a high correlation coefficient.
In order to calculate the correlation coefficient for the evaluation value of an image, a plurality of evaluation values are required for each image, but by using the evaluation value of the image by all individuals or superior individuals of each generation or multiple generations. , The correlation coefficient can be calculated. For example, the correlation matrix can be calculated by setting a predetermined generation and individual, and an image in which each correlation coefficient is equal to or more than a threshold value can be selected, and the other can not be selected.

3. 3. Reduction by clustering The evaluation values are clustered and the number of images set from each class is selected. If the number of images belonging to the class is biased, it is possible to adjust to a smaller number.
As a clustering method, a discriminant analysis method, a hierarchical cluster analysis such as Ward's method, group average method, or shortest distance method, or a non-hierarchical cluster analysis such as k-means method can be used.

（８）分離度Ｓを計算する

（９）上記（３）〜（８）を繰り返して、最小値〜最大値の範囲内の全ての閾値Ｔに対して分離度Ｓを計算する
（１０）分離度Ｓが最大となる閾値Ｔを最終的に用いる閾値とする
（１１）各クラスから設定した削減率になるように学習に用いる画像の選択を行う The discriminant analysis method for classifying into two can be carried out by the following procedure. In the examples described later, the number of learning images was reduced according to this discriminant analysis method.
(1) Calculate the histogram of the evaluation value (2) Calculate the minimum value, maximum value, and average value from the histogram (3) Select a certain threshold value T in the range from the minimum value to the maximum value (4) Set the threshold value T (5) Calculate the variance, average value, and number of pixels of class 1 (6) Calculate the variance, average value, and number of pixels of class 2 (7) Dispersion within the class, class Calculate the variance between

(8) Calculate the degree of separation S

(9) Repeat the above steps (3) to (8) to calculate the degree of separation S for all threshold values T within the range of the minimum value to the maximum value. (10) Set the threshold value T at which the degree of separation S is maximum. Set as the final threshold value (11) Select the image to be used for learning so that the reduction rate set from each class is obtained.

In the learning image reduction process, the ratio of the images to be reduced can be appropriately set. For example, the reduction rates of the learning images stored in the image storage unit 14 can be set to 20%, 60%, 80%, 90%, 95%, and the like.

Further, the learning image reduction unit 22 sets a flag or the like for identifying the image to be reduced from the learning image stored in the image storage unit 14 in the image storage unit 14. Then, the image processing unit 17 and the evaluation value calculation unit 19 can execute image processing and evaluation value calculation for only the learning image selected in the image storage unit 14 based on the flag. ..

Next, the processing procedure by the image processing apparatus of this embodiment will be described with reference to FIGS. 3 to 7. FIG. 3 is a flowchart showing a processing procedure by the image processing apparatus of the present embodiment. FIG. 4 is an explanatory diagram of evaluation value data in the evaluation result storage unit in the image processing apparatus of the present embodiment, and FIG. 5 is an explanatory diagram of ranking data in the evaluation result storage unit. FIG. 6 is a flowchart showing a processing procedure of gene manipulation in the processing procedure by the image processing apparatus. FIG. 7 is an explanatory diagram of evaluation value data of a plurality of generations in the evaluation result storage unit.

First, the image input unit 13 in the image processing device 1 inputs an image and stores it in the image storage unit 14 (step 10). Further, the teacher information input unit 15 inputs a teacher image as teacher information and stores it in the teacher information storage unit 16 (step 11).
As a specific example, it is assumed that 150 learning images and 150 verification images are stored in the image storage unit 14, and 150 teacher images corresponding to these are stored in the teacher information storage unit 16.

Next, the individual generation unit 11 generates an initial population (step 12).
Specifically, the individual generation unit 11 can generate individual information by randomly selecting and arranging nodes from the functional unit storage unit 10, and can store each identification information in the population storage unit 12.
As a specific example, it is assumed that 100 individual information is generated as an initial population.
The execution order of each step of steps 10 to 12 may be changed.

At this time, the individual generation unit 11 can store a population composed of a plurality of individual information as one generation, and store the population in the population storage unit 12 for each generation number.
A new generation population can be created by genetically manipulating the individual information stored in the population storage unit 12 by the gene manipulation unit 21. That is, a new generation population can be created by selecting or changing individual information by the genetic manipulation unit 21 and updating the population storage unit 12.
In this way, the generations in the population storage unit 12 can be sequentially added according to the execution of the generation loop shown in step 13.

The generation loop is executed for a preset number of generations. If the end condition is satisfied by the end determination described later, the generation loop may end before the execution for the set number of generations.
As a specific example, it is assumed that the number of generations of the generation loop is set to 5000 generations.

In the generation loop, the learning image reduction unit 22 determines whether or not to execute the learning image reduction process (step 14).
This criterion can be set as appropriate. For example, a predetermined set value can be set, and the learning image reduction process can be executed for all generations after the set value. It is also possible to execute reduction processing for each fixed number of generations.

As a specific example, the set value is set to 10, and after the loop of the 1st to 10th generations is completed, the learning image reduction processing is executed for each generation after the 11th generation.
In this case, in the first loop of the first generation, the learning image reduction unit 22 determines that the learning image reduction process is not executed (NO in step 14).

Next, a population loop is executed for all individual information of the population of the generation in the population storage unit 12 (step 16).
In a specific example, in the first population loop, steps 17 and 18 are repeatedly executed for each of the 100 individual pieces of information.

That is, the image processing unit 17 executes image processing for each individual information of the population of the generation in the population storage unit 12 (step 17).
In a specific example, the image processing unit 17 sequentially executes a plurality of functional units constituting individual information for each of the 150 learning images, and an image or the like as output information corresponding to each learning image. Can be created and stored in the output information storage unit 18.

Next, the evaluation value calculation unit 19 calculates the evaluation value by comparing the output information with the teacher information for each individual information and the learning image of the population of the generation in the output information storage unit 18 (step). 18).
At this time, the evaluation value calculation unit 19 can calculate the mean square error as the evaluation value by using the output information and the teacher information, for example, by the least squares method.
Then, the evaluation value calculation unit 19 can store the obtained evaluation value in the evaluation result storage unit 20 for each individual information and each learning image.

FIG. 4 shows an example of evaluation value data stored in the evaluation result storage unit 20. As shown in the figure, the evaluation value calculated by the evaluation value calculation unit 19 can be stored for each individual information and each learning image for the generation.
The evaluation value calculation unit 19 can also calculate and store the statistical value of the evaluation value for each individual information. As the statistical value of the evaluation value, for example, an average value or a standard deviation can be used.

Next, the evaluation value calculation unit 19 ranks individual information based on the evaluation value stored in the evaluation result storage unit 20, and stores the result in the evaluation result storage unit 20 (step 19).
FIG. 5 shows an example of ranking data stored in the evaluation result storage unit 20. As shown in the figure, the ranking of individual information can be performed based on the statistical value of the evaluation value or the like.

In addition, the evaluation value calculation unit 19 performs a determination process of whether or not the end condition is satisfied (step 20). If the end condition is satisfied, the process by the image processing device 1 is terminated.
If the termination condition is not satisfied, the gene manipulation unit 21 executes the gene manipulation process (step 21).
As a result, the genetic engineering unit 21 can additionally update the population of a new generation to the population storage unit 12.

Specifically, as shown in FIG. 6, the elite individual preservation unit in the gene manipulation unit 21 selects the individual information having the highest evaluation value as the individual information of the next-generation population (step 30).
Further, the individual selection unit in the gene manipulation unit 21 selects as individual information of the next-generation population with a higher probability as the individual information having a higher evaluation value ranks (step 31).

Further, the crossover portion in the gene manipulation unit 21 executes a crossover process on the individual information selected by the individual selection unit to generate new individual information (step 32).
In addition, the mutation unit in the gene manipulation unit 21 performs mutation processing or the like based on the individual information selected by the individual selection unit or without using these individual information to generate new individual information ( Step 33).
The order of each process by the elite individual preservation unit, the individual selection unit, the crossover unit, and the mutation unit may be exchanged within a range in which each process can be executed.

By this genetic manipulation process, individual information of the second generation population can be added to the population storage unit 12.
By repeating the generation loop in this way, the individual information of the new generation population can be sequentially added to the population storage unit 12.
Then, when the 10th generation generation loop ends, as shown in FIG. 7, the evaluation result storage unit 20 stores evaluation value data for 10 generations for each individual information and for each learning image.

Next, the 11th generation generation loop is executed.
In the specific example, since the set value is 10, the learning image reduction unit 22 determines that the learning image reduction process is executed (YES in step 14), and executes the learning image reduction process (step 15). ).
That is, the learning image reduction unit 22 reduces the learning images image-processed by the image processing unit 17 based on the evaluation value of each learning image stored in the evaluation result storage unit 20.
Specifically, the learning image reduction unit 22 applies to the above-mentioned "1. Reduction by ranking of evaluation values", "2. Reduction using correlation coefficient of evaluation values", or "3. Reduction by clustering". Therefore, the number of learning images can be reduced.

At this time, the learning image reduction unit 22 can use the evaluation value of the best individual information of the generation as the “evaluation value of each learning image”. In addition, ranking information based on statistical values for the evaluation values of all individuals of the generation can be used, or statistical values of predetermined higher evaluation values can be used. Further, it is also possible to use statistical values of evaluation values of a plurality of generations (for example, the average value of evaluation values of the best individual information for 10 generations).
When using evaluation values between generations, the range of evaluation values (minimum value to maximum value) is different, so it is possible to calculate the ranking after normalizing the evaluation values with minimum value 0, maximum value 1, etc. can.

In the learning image reduction process, the ratio of the images to be reduced can be appropriately set. For example, the reduction rates of the learning images stored in the image storage unit 14 can be set to 20%, 60%, 80%, 90%, 95%, and the like.
As a specific example, when the reduction rate is 80%, 30 images are selected from 150 learning images, and 30 learning images are used in the population loop in the generation loop, and these are selected. Image processing and evaluation value calculation are repeated only for the learning image.

At this time, the learning image reduction unit 22 sets a flag in the image storage unit 14 for identifying the image to be reduced from the learning image stored in the image storage unit 14. Then, the image processing unit 17 and the evaluation value calculation unit 19 can execute image processing and evaluation value calculation for only the learning image selected in the image storage unit 14 based on the flag. ..

The learning image reduction unit 22 selects a learning image in which image processing and calculation of the evaluation value are executed based on the evaluation value of each learning image stored in the evaluation result storage unit 20, and this evaluation. Values vary in the population loop.
That is, in the image processing system of the present embodiment, the evaluation value of the learning image changes as the learning progresses, and the learning image reduction unit 22 reduces the learning image based on the evaluation value, and thus is used for learning. The learning image to be used is not fixed, but changes according to the image processing algorithm (individual information) that evolves due to the generation loop.

Further, with respect to the 150 verification images stored in the image storage unit 14 shown in the specific example, the image processing unit 17 executes image processing based on specific individual information at a predetermined timing, and for each verification. Generate output information corresponding to the image. Then, the evaluation value calculation unit 19 compares the output information corresponding to each verification image with the teacher information, calculates the evaluation value for each specific individual information and for each verification image, and causes the evaluation result storage unit 20 to calculate the evaluation value. Remember.

The predetermined timing may be, for example, for each generation. Specifically, after ranking the individual information in step 19, image processing for each verification image can be executed using the best individual information in each generation, and output information corresponding to each can be generated. ..
Further, the predetermined timing may be set for each set generation (for example, every 10 generations), and after ranking the individual information in step 19, the best individual information can be used in the same manner.

Furthermore, the above-mentioned predetermined timing is set to the timing when the evaluation value of the best individual information is improved (the timing when the evaluation value using the learning image is improved) and the final generation (after the generation loop is completed (evolutionary computation is completed). The same can be done using the individual information finally obtained as (later)). In this way, the timing for processing the verification image can be arbitrarily set.
Also in the second embodiment, the processing for the verification image can be performed at the same timing.

According to the image processing system of the present embodiment as described above, the image used in the training is obtained by using the evaluation value calculated by comparing the output image obtained by performing the image processing on the training image with the teacher image. It can be reduced and the learning time can be shortened while maintaining the performance of the generated image processing algorithm.

Further, since the evaluation value calculated at the time of learning is used in the reduction of the learning image, it is not necessary to set the feature amount for the reduction of the learning image at the initial stage. Further, since the evaluation value for the image changes as the learning progresses, the image used for the learning is not fixed. Furthermore, since the evaluation value represents the characteristics of the image processing algorithm generated by evolutionary computation, it is possible to expect an improvement in the performance of the generated image processing algorithm.

[Second Embodiment]
Next, the image processing system and the image processing program according to the second embodiment of the present invention will be described with reference to FIG. FIG. 8 is a block diagram showing a configuration of an image processing device corresponding to the image processing system of the present embodiment.

As shown in FIG. 8, the image processing apparatus 1a of the present embodiment includes a functional unit storage unit 10a, an individual generation unit 11a, a population storage unit 12a, an image input unit 13a, an image storage unit 14a, and a teacher information input unit 15a. It includes a teacher information storage unit 16a, an image processing unit 17a, an output information storage unit 18a, an evaluation value calculation unit 19a, an evaluation result storage unit 20a, a gene manipulation unit 21a, and a learning image reduction unit 22a. Further, the population storage unit 12a includes a parent population storage unit 121a and a child population storage unit 122a.

The functional unit storage unit 10a, the image input unit 13a, the image storage unit 14a, the teacher information input unit 15a, the teacher information storage unit 16a, the output information storage unit 18a, and the evaluation result storage unit 20a in the image processing device 1a of the present embodiment are , The same as in the first embodiment. Further, other configurations in the image processing apparatus 1a of the present embodiment can be the same as those of the first embodiment except for the points described below.

The individual generation unit 11a generates individual information composed of an array of a plurality of nodes by using the nodes stored in the functional unit storage unit 10a. At this time, the individual generation unit 11a can define the execution order of a plurality of nodes in a network shape that can include a closed cycle in the individual information, as in the first embodiment.

The individual generation unit 11a randomly generates a plurality of parent individual information as individual information, and stores an initial parent population composed of these parent individual information in the parent population storage unit 121a. The parent population can be stored in the parent population storage unit 121a for each generation number.

The population storage unit 12a includes a parent population storage unit 121a and a child population storage unit 122a.
The parent population storage unit 121a stores a parent population consisting of a plurality of individual information generated by the individual generation unit 11a. In addition, the parent population of one generation or a plurality of generations can be stored, and each individual information can be stored for each generation number.
The child population storage unit 122a stores a child population composed of a plurality of individual information generated by the gene manipulation unit 21a. In addition, child populations of one generation or a plurality of generations corresponding to the parent population can be stored, and each individual information can be stored for each generation number.

In the present embodiment, a set of a parent population and a child population is included in one generation, and these sets are generated for a plurality of generations and stored in the parent population storage unit 121a and the child population storage unit 122a, respectively. Will be done.
The evaluation value calculation unit 19a, which will be described later, can rank the child individual information for each generation or for each set generation, and store the ranking information in the evaluation result storage unit 20a.

The image processing unit 17a inputs an image from the image storage unit 14a, inputs child individual information from the child population storage unit 122a, and sequentially executes a plurality of nodes included in the child individual information based on the child individual information. do. Then, the image processing unit 17a can create an image as output information for each of the learning image and the verification image, and store this output information in the output information storage unit 18a for each child individual information and each image. ..

The evaluation value calculation unit 19a inputs output information from the output information storage unit 18a, and also inputs teacher information corresponding to the output information from the teacher information storage unit 16a.
Then, the evaluation value is calculated by comparing the output information and the teacher information, and the obtained evaluation value can be stored in the evaluation result storage unit 20a for each child individual information and each image.

Further, the evaluation value calculation unit 19a completes the calculation of the evaluation value for all the child individual information (or the child individual information for one generation) stored in the child population storage unit 122a in the population storage unit 12a and evaluates. The child individual information can be ranked based on the evaluation values at the timing stored in the result storage unit 20a, and the obtained ranking information can be stored in the evaluation result storage unit 20a.
Further, the evaluation value calculation unit 19a can perform an end determination for determining whether or not to end the image processing by the image processing unit 17a.

The gene manipulation unit 21a executes a gene manipulation on the parent individual information in the parent population storage unit 121a, generates a child population consisting of a plurality of individual information corresponding to the parent population, and generates a child population storage unit. It can be stored in 122a.
Specifically, the gene manipulation unit 21a first randomly selects parent individual information for generating child individual information from the parent population in the parent population storage unit 121a. Next, a child population consisting of the same child individual information as the selected parent individual information is stored in the child population storage unit 122a. Further, a crossing process is executed on the selected parent individual information to generate a plurality of child individual information, and the child population composed of the child individual information is stored in the child population storage unit 122a.

In addition, the gene manipulation unit 21a executes mutation processing on the selected parent individual information using the node stored in the functional unit storage unit 10a, and has a certain probability that a part of the parent individual information or a part of the parent individual information or It is also possible to store the child population consisting of the child individual information obtained by rewriting all of them in the child population storage unit 122a. Furthermore, as a mutation process, new individual information can be generated using the node stored in the functional unit storage unit 10a, and some or all of the parameters of the node of the selected parent individual information can be randomly selected. It is also possible to mutate the parameters by rewriting.

In this way, the gene manipulation unit 21a generates a child population consisting of a plurality of individual information by executing the gene manipulation on the parent individual information in the parent population storage unit 121a, and these are used as the child population. It is stored in the storage unit 122a.

Further, the gene manipulation unit 21a randomly selects the child individual information having the best evaluation value and the child individual information selected with the probability based on the ranking information of the child individual information in the evaluation result storage unit 20a from the parent population. The parent individual information in the parent population storage unit 121a can be updated by replacing the parent individual information selected in.
As a result, the gene manipulation unit 21a can add new individual information of the parent population to the parent population storage unit 121a in the population storage unit 12a.
Then, by repeating the generation loop described later, the individual information of the parent population of a new generation can be sequentially added in the population storage unit 12a.

The learning image reduction unit 22a determines whether or not to execute the learning image reduction process.
Further, when the learning image reduction unit 22a determines that the reduction process is to be executed, the learning is image-processed by the image processing unit 17a based on the evaluation value of each learning image stored in the evaluation result storage unit 20a. Reduce images for use.
In the learning image reduction process, the learning image reduction unit 22a can reduce the learning image by various methods similar to those in the first embodiment based on the evaluation value of each learning image.

In the learning image reduction process, the ratio of the images to be reduced can be appropriately set. For example, the reduction rates of the learning image stored in the image storage unit 14a can be set to 20%, 60%, 80%, 90%, 95%, and the like.

Further, the learning image reduction unit 22a sets a flag or the like for identifying the image to be reduced from the learning image stored in the image storage unit 14a in the image storage unit 14a. Then, the image processing unit 17a and the evaluation value calculation unit 19a can execute image processing and evaluation value calculation for only the selected learning image in the image storage unit 14a based on the flag. ..

Next, the processing procedure of the image processing apparatus corresponding to the image processing system of the present embodiment will be described with reference to FIG. FIG. 9 is a flowchart showing a processing procedure by the image processing apparatus corresponding to the image processing system of the present embodiment.

First, the image input unit 13a in the image processing device 1a inputs an image and stores it in the image storage unit 14a (step 40). Further, the teacher information input unit 15a inputs a teacher image as teacher information and stores it in the teacher information storage unit 16a (step 41).
Next, the individual generation unit 11a generates an initial parent population (step 42).
The execution order of each step of steps 40 to 42 may be changed.

At this time, the individual generation unit 11a generates parent individual information by randomly selecting and arranging nodes from the functional unit storage unit 10a, and the parent individual group storage unit 121a in the population storage unit 12a for each identification information. Can be memorized in.
The parent population in the parent population storage unit 121a is updated by the update processing of the parent population described later, and the following processing is repeatedly performed for the set generation in the generation loop (step 43).

In the generation loop, the learning image reduction unit 22a determines whether or not to execute the learning image reduction process (step 44).
This criterion can be set as appropriate. For example, a predetermined set value can be set, and the learning image reduction process can be executed for all generations after the set value. It is also possible to execute reduction processing for each fixed number of generations.

When the learning image reduction unit 22a determines that the learning image reduction process is to be executed (YES in step 44), the learning image reduction unit 22a executes the learning image reduction process (step 45).
That is, the learning image reduction unit 22a reduces the learning images image-processed by the image processing unit 17a based on the evaluation value of each learning image stored in the evaluation result storage unit 20a.

Specifically, the learning image reduction unit 22a is set to the above-mentioned "1. Reduction by ranking of evaluation values", "2. Reduction using correlation coefficient of evaluation values", or "3. Reduction by clustering". Therefore, the number of learning images can be reduced.
Further, when the learning image reduction unit 22a determines that the learning image reduction process is not executed (NO in step 44), the learning image reduction unit 22a executes step 46 without executing the learning image reduction process.

Next, the genetic manipulation unit 21a randomly selects parent individual information for generating child individual information from the parent population in the parent population storage unit 121a. (Step 46). Then, genetic manipulation is executed on the selected parent individual information (step 47). The selection of parent individual information in step 46 may be performed as part of the genetic manipulation in step 47.
Then, the gene manipulation unit 21a executes crossover processing or mutation processing on the randomly selected parent individual information to generate a plurality of child individual information, and the child population consisting of these child individual information is generated as a child. It is stored in the population storage unit 122a.

Next, in the child population loop (step 48), the image processing unit 17a executes image processing on all the child individual information of the generation in the child population storage unit 122a (step 49).
At this time, the image processing unit 17a can create an image or the like as output information and store it in the output information storage unit 18a for each child individual information and each image. At this time, when the detection target exists in the learning image and the verification image, the image processing unit 17a creates an image to which the information indicating the region where the detection target exists is added and stores it in the output information storage unit 18a. Can be made to.

Next, the evaluation value calculation unit 19a calculates the evaluation value by comparing the output information and the teacher information (step 50), and stores the obtained evaluation value in the evaluation result storage unit 20a for each child individual information and for each image. Let me.
Then, for all the child individual information in the child population storage unit 122a, the processing from the image processing to the calculation and storage of the evaluation value is repeatedly executed.

Next, the evaluation value calculation unit 19a ranks the child individual information based on the evaluation value stored in the evaluation result storage unit 20a, and stores the result in the evaluation result storage unit 20a (step 51). .. The ranking of individual information can be performed based on the statistical value of the evaluation value or the like.

Then, the gene manipulation unit 21a selects the child individual information having the best evaluation value and the child individual information selected with the probability based on the ranking information of the child individual information in the evaluation result storage unit 20a, and the parent individual in step 46. The parent individual information of the parent population storage unit 121a in the population storage unit 12a can be updated by replacing the parent individual information randomly selected from the group with the above parent individual information (step 52).

By such processing of the gene manipulation unit 21a, individual information of the parent population as a new generation can be added to the parent population storage unit 121a in the population storage unit 12a.
Then, by repeating the generation loop, the individual information of the parent population of a new generation can be sequentially added in the population storage unit 12a.

Further, the evaluation value calculation unit 19a performs a determination process of whether or not the end condition is satisfied (step 53), and if the end condition is satisfied, the process by the image processing device 1a ends.

When the processing by the image processing device 1a is terminated, the evaluation value calculation unit 19a calculates the evaluation value for all the individual information in the final generation, ranks the individual information based on the evaluation value, and obtains the ranking information. Can be stored in the evaluation result storage unit 20a.

According to such an image processing apparatus of the present embodiment, gene manipulation is executed when a parent is selected from a parent population to generate child individual information, and specific child individual information is transmitted to the parent individual in the parent population. By replacing the information, the parent population can be updated and individual information of a new generation can be added. This makes it possible to search for optimal individual information while maintaining more evolutionary diversity than in the first embodiment.
In addition, it is possible to maintain the performance of the generated image processing algorithm and shorten the learning time in a state where such evolutionary diversity is further maintained.

The image processing apparatus of the above embodiment can be realized by using a computer controlled by the image processing program of the present invention. The CPU of the computer sends a command to each component of the computer based on the image processing program, and predetermined processing required for the operation of the image processing apparatus, for example, individual generation processing, image reduction processing for learning, image processing, and evaluation. Have them perform value calculation processing, gene manipulation processing, etc. As described above, each process and operation in the image processing apparatus of the present invention can be realized by specific means in which the program and the computer cooperate.

The program is stored in a recording medium such as ROM or RAM in advance, and the program is read by the computer from the recording medium mounted on the computer and executed. However, the program can also be read by the computer via a communication line, for example.
Further, the recording medium for storing the program can be configured by, for example, a semiconductor memory, a magnetic disk, an optical disk, or any other recording means that can be read by any computer.

As described above, according to the image processing system and the image processing program according to the embodiment of the present invention, the evaluation calculated by comparing the output image obtained by performing image processing on the learning image with the teacher image. By making it possible to select an image to be used in training using a value, it is possible to shorten the training time while maintaining the performance of the generated image processing algorithm.
In addition, since the evaluation value for the image changes as the learning progresses, the image used for training is not fixed, and the evaluation value represents the characteristics of the image processing algorithm generated by evolutionary computation. It is also possible to expect an improvement in the performance of the processing algorithm.

Hereinafter, a test in which image processing for detecting a detection target from an image is optimized using the image processing system according to the embodiment of the present invention will be described with reference to FIGS. 10A, B to 15.
As an example, an image processing apparatus corresponding to the image processing system according to the second embodiment of the present invention was used.
Further, as a reference example, a modified image processing apparatus of the embodiment was used. Specifically, in Reference Example 1, the learning image reduction unit changes the learning image to be randomly selected at a fixed ratio prior to the generation loop, and in Reference Example 2, the learning image reduction unit processes the image. Was not executed, and the reduction rate of the learning image was set to 0%.

First, the images used in this test were created programmatically based on the following procedure and set values. As sample images, sample image 1 (both detection target and non-detection target are not drawn), sample image 2 (detection target is drawn), and sample image 3 (non-detection target is drawn). 200 types were created for each. Then, out of each of the 200 images, 50 images were used as learning images (150 images in total) and 150 images were used as verification images (450 images in total). It should be noted that the classification of which to use as the learning image and the verification image was randomly performed.

(Procedure 1) Creating a background image
An image of 256 pixels x 256 pixels, 8 bits, and 0 brightness (black image) was prepared, and 10 circles with a diameter of 22 to 24, a width of 3 to 5, and a brightness of 71 to 149 were drawn on this image. Then, brightness 0 to 20 was randomly determined and added to each pixel of the obtained image to obtain a background image.

(Procedure 2) Creation of a copy image of 600 background images An arbitrary position of the background image was copied to an 8-bit image of 64 pixels × 64 pixels to create a copy image of 600 background images.
The background images in the learning image and the verification image used in the examples and the reference examples are shown in FIG. 10A, and the copy image at an arbitrary position of the background image is shown in FIG. 10B.

(Procedure 3) Creation of sample image 1 (similar to the background image, both the detection target and the non-detection target are not drawn)
The copy images of images Nos. 1 to 200 created in step 2 were designated as sample images 1. In addition, teacher images corresponding to each image were created. Since the detection target is not drawn in the sample image 1, each teacher image is black.
FIG. 11A shows a sample image 1 as a learning image or a verification image used in Examples and Reference Examples.

(Procedure 4) Creation of sample image 2 (image in which the detection target is drawn)
The detection target was drawn on the copy images of images Nos. 201 to 400 created in step 2 to obtain sample images 2. At this time, one straight line having a thickness of 2 to 4 pixels, a border of 0 to 2 pixels, a brightness of 125 to 275, and a length of 32 to 64 pixels was drawn at an arbitrary position of each image, and this was set as a detection target. In addition, the teacher image corresponding to each image was created by painting the part corresponding to the detection target in white.
FIG. 11B shows a sample image 2 as a learning image or a verification image used in Examples and Reference Examples.

(Procedure 5) Creation of sample image 3 (image in which a non-detection target is drawn)
The non-detection target was drawn on the copy images of images Nos. 401 to 600 created in step 2 to obtain sample image 3. At this time, a circle with a diameter of 8 to 16 and a brightness of 155 to 205 was drawn at an arbitrary position of each image, and this was set as a non-detection target. Since the detection target is not drawn in the sample image 3 (the non-detection target is drawn), each teacher image is black.
FIG. 11C shows a sample image 3 as a learning image or a verification image used in Examples and Reference Examples.

Evolutionary computation in Examples and Reference Examples was performed by MGG (Minimal Generation Gap), which is the second embodiment, using various parameters shown in FIG. As the images, as described above, 150 learning images and 450 verification images of 64 pixels × 64 pixels were used, respectively. The network has 1 input node, 1 output node, and 7 processing nodes. Evolutionary computation shows that the number of generations is 5000, the number of parent individuals is 100, the number of offspring is 62 (2 elites, 20 uniform crossovers, 30 mutations, 10 parameter mutations), and the rate of change is 1. The crossover was 20%, the mutation was 20%, and the parameter variation was 20%.

In the parent selection from the parent population, two parent individual information was randomly selected from the 100 parent individual information in the initial parent population generated by the individual generation unit and deleted from the parent population.
Then, genetic manipulation was performed on these parent individual information to generate 62 child individual information (same as parent: 2, uniform crossover: 20, mutant individual: 30 individual information, parameter variation). 10).

In addition, after the image processing is executed, the child individual information having the best evaluation value is selected based on the evaluation result of the child individual information, and the total of two child individual information selected with probability is the parent individual information. The parent population was renewed by returning to the parent population.

The nodes shown in FIGS. 13A to 13I were used to execute the image processing. The node shown in FIG. 13A performs image processing (noise removal), the node shown in FIG. 13B performs image processing (brightness correction), and the node shown in FIG. 13C performs image processing (binary value). The node shown in FIG. 13D performs image processing (edge detection), and the node shown in FIG. 13E performs image processing (frequency filter), which is shown in FIG. 13F. The node performs image processing (calculation), the node shown in FIG. 13G performs image processing (morphology), and the node shown in FIG. 13H performs image processing (detection target determination). Yes, the node shown in FIG. 13I performs image processing (branching).

The evaluation value was calculated by the following formula using the mean square error in both the examples and the reference examples. In this evaluation value, the lower the value, the better the evaluation.

Here, for the reduction of the learning image, five types of reduction rates (20%, 60%, 80%, 90%, 95%) were set in the examples, and tests were performed for each of them.
Also, in Reference Example 1, five types of reduction rates (20%, 60%, 80%, 90%, 95%) are set, and the images for learning are randomly reduced prior to the execution of the generation loop for learning. Tests were conducted for each of them in the same manner as in Examples except for the reduction of images.
Further, in Reference Example 2, the reduction rate was set to 0%, and the test was conducted in the same manner as in Example except for the reduction of the learning image without reducing the learning image.

In the reduction of learning images, the reduction process is not performed in the first 100 generations, and the evaluation value of each learning image calculated based on the individual information having the highest evaluation value for each 100 generations after the 100th generation is used for discriminant analysis. The learning image to be used was selected by the analysis method.
At this time, the learning images were divided into two classes based on the evaluation value by the discriminant analysis method, and the learning images were selected so as to have the set reduction rate and the same number of learning images selected from each class.
If there are no images in one class, select evenly for the rank of the evaluation value, and if there are few images in one class and the set reduction rate is not satisfied, do not perform any operation. I learned as it was.
The execution of image processing using the verification image and the calculation of the evaluation value were performed using the individual information having the best evaluation value at the timing when the learning using the learning image was completed.

Then, with respect to the examples and the reference examples, the above processing was repeatedly executed for 5000 generations, and the processing was completed. Further, these were repeated three times, and the average value of the evaluation value of the individual information having the highest evaluation value and the time required for learning of 5000 generations was calculated.
The results are shown in FIGS. 14A, 14B, and 15. FIG. 14A shows the learning time, the learning result (evaluation value), and the verification result (evaluation) according to the examples (reduction rate 20%, 60%, 80%, 90%, 95%) and the reference example 2 (reduction rate 0%). FIG. 14B is a diagram showing a graph showing the value), and FIG. 14B shows the learning time according to Reference Example 1 (reduction rate 20%, 60%, 80%, 90%, 95%) and Reference Example 2 (reduction rate 0%). It is a figure which shows the graph which shows the learning result (evaluation value), and the verification result (evaluation value). Further, FIG. 15 is a diagram showing a graph comparing the learning results (evaluation values) and the verification results (evaluation values) according to the examples and the reference examples.

In these figures, the left vertical axis is the time [min (minutes)] required for learning of the 5000 generations. The right vertical axis is an evaluation value calculated using the best individual information after learning, which was obtained for each of the learning image and the verification image. In addition, each plot at 0% on the horizontal axis is a case where all the training images are selected (reduction rate 0%), and shows the result of Reference Example 2.

From the results of FIGS. 14A and 14B, the learning time can be significantly shortened in the example and the reference example 1 in which a part of the learning image is reduced as compared with the reference example 2 in which the learning image is not reduced.
Further, when the Examples and the Reference Example 1 are compared, those having the same selection rate of the learning images have substantially the same learning time.

Next, regarding the performance of the generated image processing algorithm, in the example, when the reduction rate is 80% for both the learning result and the verification result, the same performance as that of the reference example 2 is obtained. Further, when the reduction rate is 60%, the best performance is obtained, which is better than that of Reference Example 2.
On the other hand, in Reference Example 1, the learning result shows the same performance as in Reference Example 2 with a reduction rate of 60%, but in the verification result, even if the reduction rate is 20%, it is equivalent to Reference Example 2. Performance has not been obtained.
For this reason, if the learning image to be reduced is randomly selected as in Reference Example 1 instead of using the evaluation value as in the example, the performance of the image processing algorithm is maintained and the time is shortened. It was suggested that it was difficult.

Further, in the examples, better performance is shown in both the learning result and the verification result as compared with the reference example 1. In particular, the difference is remarkable in the verification results of the reduction rates of 60% and 80%. It is presumed that the reason is that an image that can be learned efficiently is selected by reducing the image for learning by using the evaluation value.

Further, in the examples, the performance was further improved at a part of the selection rate (learning result and verification result with a reduction rate of 60%) as compared with Reference Example 2 in which the learning image was not reduced. The cause is that, in the embodiment, an image processing algorithm specialized for the selected image is generated by selecting and learning the image, and the selected image changes, so that the image processing algorithm has various characteristics. Is generated, and it is presumed that the performance was improved by crossing them.

As described above, by using the image processing system and the image inspection program of the present embodiment, it was possible to obtain an image processing algorithm with maintained performance in a learning time shortened by about 77%.
In addition, a better performing image processing algorithm could be obtained with a learning time reduced by about 63%.
In this way, by selecting the learning image to be used in the image processing of each generation based on the evaluation value of the learning image, the learning time is significantly reduced while maintaining or improving the performance of the image processing algorithm. I found that I could do it.

It goes without saying that the present invention is not limited to the above embodiments, and various modifications can be made within the scope of the present invention.
For example, each configuration in the image processing device can be distributed to a plurality of information processing devices, or can be used as a node including a node other than that used in the embodiment.
Further, as a method of selecting an image to be reduced by using the evaluation value, it is possible to appropriately change a method of reducing evenly, a method of changing the probability of reduction by using a sigmoid function, and the like.

Further, the above image processing device is used for image inspection, the image input unit inputs an image for inspection, and the image processing unit processes the image for inspection based on the individual information having the highest evaluation value. It is also possible to have a configuration including a processing result storage unit that stores the processing result information obtained as a result.

INDUSTRIAL APPLICABILITY The present invention is a case where inspection or the like is performed based on image data of equipment or a product, and can be suitably used for obtaining an information processing apparatus for image inspection with optimized image processing. ..

1,1a Image processing device 10,10a Functional unit storage unit 11,11a Individual generation unit 12, 12a Individual group storage unit 121a Parent population storage unit 122a Child population storage unit 13, 13a Image input unit 14, 14a Image storage unit 15,15a Teacher information input unit 16,16a Teacher information storage unit 17,17a Image judgment unit 18,18a Output information storage unit 19, 19a Evaluation value calculation unit 20, 20a Evaluation result storage unit 21,21a Gene manipulation unit 22, 22a Image reduction department for learning

Claims (10)

An image processing system that optimizes image processing using genetic manipulation, which is an optimization method that mathematically simulates the evolution of living organisms.
A population storage unit that stores multiple individual information that defines the execution order of multiple functional units in a linear structure, tree structure, or network.
An image storage unit that stores multiple learning images,
An image processing unit that processes the learning image based on the individual information and generates output information corresponding to the learning image.
A teacher information storage unit that stores teacher information corresponding to the learning image for evaluating the output information, and a teacher information storage unit.
An evaluation value calculation unit that compares the output information with the teacher information, calculates an evaluation value for each individual information and each learning image, and ranks the individual information based on the evaluation value for each individual information. ,
A genetic manipulation unit that updates the population storage unit by selecting or changing the individual information based on the ranking information of the individual information.
An image processing system including a learning image reduction unit that reduces learning images processed by the image processing unit based on an evaluation value for each learning image. The genetic engineering unit updates the population memory unit to create information on a plurality of individuals of a new generation.
When the image processing unit creates new generation individual information by the genetic manipulation unit, the new generation individual information is used as the individual information.
After the set generation or for each set generation, the learning image reduction unit is a part of the learning image stored in the image storage unit based on the evaluation value for each learning image. Select an image and
The image processing system according to claim 1, wherein the image processing unit uses the selected partial image as the learning image. The learning image reduction unit clusters the learning images based on the evaluation value of the learning images after the set generation or for each set generation, and a part of the learning images from each class. The image processing system according to claim 1 or 2, wherein the image processing system is selected. The image processing system according to claim 3, wherein the image reduction unit for learning classifies the images for learning by using a discriminant analysis method, a hierarchical cluster analysis, or a non-hierarchical cluster analysis. The learning image reduction unit selects a part of the learning images after the set generation or for each set generation based on the rank or the correlation coefficient of the evaluation value of the learning image. The image processing system according to claim 1 or 2, wherein the image processing system is characterized. The image storage unit stores a plurality of verification images together with the learning image.
The teacher information storage unit stores the teacher information corresponding to the verification image for evaluating the output information, and stores the teacher information.
The image processing unit processes the verification image based on specific individual information at a predetermined timing, and generates output information corresponding to the verification image.
A claim characterized in that the evaluation value calculation unit compares the output information corresponding to the verification image with the teacher information and calculates an evaluation value for each specific individual information and for each verification image. Item 4. The image processing system according to any one of Items 1 to 5. An inspection image storage unit that stores inspection images,
Claim 1 is further provided with a processing result storage unit that stores the processing result information obtained by processing the image for inspection by the image processing unit based on the individual information having the highest evaluation value. The image processing system according to any one of 6 to 6. An image processing system that optimizes image processing using genetic manipulation, which is an optimization method that mathematically simulates the evolution of living organisms.
Computer,
A population storage unit that stores multiple individual information that defines the execution order of multiple functional units in a linear structure, tree structure, or network.
Image storage unit that stores multiple learning images,
An image processing unit that processes the learning image based on the individual information and generates output information corresponding to the learning image.
A teacher information storage unit that stores teacher information corresponding to the learning image for evaluating the output information,
An evaluation value calculation unit that compares the output information with the teacher information, calculates an evaluation value for each individual information and each learning image, and ranks the individual information based on the evaluation value for each individual information.
A genetic manipulation unit that updates the population storage unit by selecting or changing the individual information based on the ranking information of the individual information, and
An image processing program characterized by functioning as a learning image reduction unit that reduces learning images processed by the image processing unit based on an evaluation value for each learning image. By having the gene manipulation unit update the population memory unit, a plurality of individual information of a new generation is created.
When the image processing unit creates new generation individual information by the genetic manipulation unit, the new generation individual information is used as the individual information.
A part of the learning image stored in the image storage unit based on the evaluation value for each learning image after the generation set in the learning image reduction unit or for each generation set. Let me select an image
The image processing program according to claim 8, wherein the image processing unit is made to use the selected partial image as the learning image. The learning image reduction unit clusters the learning image based on the evaluation value of the learning image after the set generation or for each set generation, and a part of the learning image from each class. The image processing program according to claim 8 or 9, wherein the image processing program is selected.

Images