JP6208018B2 - Image recognition algorithm combination selection device - Google Patents

Description

  Embodiments described herein relate generally to an image recognition algorithm combination selection apparatus.

  Image recognition technology is increasingly used in various fields such as pedestrian detection in in-vehicle cameras and security by face recognition.

  The algorithm for image recognition depends on the recognition target, and the effective algorithm differs depending on the target to be identified. For example, certain algorithms may be able to identify pedestrians but are not suitable for identifying fruits. In addition, there are a number of alternative algorithms that are difficult to select without proper knowledge. Therefore, conventionally, an appropriate algorithm is selected based on the knowledge of an expert.

  However, in response to the growing needs of applications related to image recognition, there is a demand for accelerating the development of such applications and making it possible to determine the optimal algorithm according to the object even if not familiar with image recognition. It is growing.

  When attempting to select an algorithm for image recognition mechanically, there is a reality that the variety of image recognition algorithms, in other words, the number of combinations of algorithms to be evaluated and their parameters is enormous. Therefore, it takes time to create a dictionary for evaluation, and there is a problem that it takes a very long time to evaluate a large number of combinations.

JP 11-259282 A JP-A-10-149280 JP 2000-56958 A

  The problem to be solved by the present invention is to provide an image recognition algorithm combination selection device capable of selecting a combination configuration of image recognition algorithms according to a recognition target even if it is not an expert.

  An image recognition algorithm combination selection apparatus according to an embodiment is an apparatus for selecting a combination of algorithm parts for image recognition, and includes learning data including image data obtained by cutting out an object to be identified from an image and an identification unit, and the image The image format information of the data is input, the learning data, the image data, and the set of the image format information are all set as a learning data list, and a part is extracted from the learning data list and used for image recognition performance evaluation. A learning data list determination unit for creating an extracted learning data list for creating a dictionary, algorithm information including a start condition representing the algorithm component and information for execution of the algorithm component, and the connection possible Dependency information representing a combination of algorithm parts and the algorithm part An image recognition algorithm determination unit that creates an algorithm candidate list using parameter set information representing all the parameters used in the above, and the parameter set information of default values and the extracted learning for the algorithm candidate list A dictionary creation unit that creates a dictionary using a data list; and a performance evaluation unit that evaluates the algorithm candidate list using the parameter set information and the dictionary for the extracted learning data list, The image recognition algorithm determination unit sets the parameters of the parameter set information with the algorithm candidate list having the best evaluation result of the performance evaluation unit as the optimal algorithm candidate list, and the dictionary creation unit The dictionary corresponding to the algorithm candidate list The performance evaluation unit re-evaluates the performance of the optimal algorithm candidate list using the re-created dictionary, and the image recognition algorithm determination unit determines the optimal parameter based on the re-evaluation result. The set is determined, and the dictionary creation unit finally recreates the dictionary using all the learning data.

It is a block diagram which shows schematic structure of the image recognition algorithm combination selection apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the information of the algorithm component stored in an algorithm component storage part. It is a figure which shows an example of the information of dependence. It is a figure which shows an example of the information of a parameter. It is a figure which shows an example of the information of learning data. It is a figure which shows an example of an algorithm candidate list. It is a figure which shows an example of an algorithm performance information list. It is a flowchart which shows the flow of a process until determination of an algorithm combination. It is a figure which shows an example of a startable algorithm information list. It is a figure which shows an example of the data produced | generated every time process 1 is called recursively. It is a figure explaining the operation | work in the process 1, and is a figure explaining calculation of the lower stage data in the dotted line of FIG. It is a figure which shows an example of the learning data list after extraction. It is a figure which shows an example of calculation of an evaluation result. It is a figure which shows an example of the optimal algorithm candidate list | wrist obtained by the process to step S811. It is a flowchart which shows the flow of a process until determination of a parameter combination. It is a figure which shows an example of the data produced | generated every time process 3 is called recursively. It is a figure explaining the operation | work in the process 3. FIG. It is a figure which shows an example of the optimal algorithm candidate list | wrist obtained by the process to step S1505. It is a flowchart which shows the flow of a process until an algorithm candidate list | wrist and a dictionary output. It is a figure explaining the dictionary created using all the learning data.

  Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same portions are denoted by the same reference numerals, and redundant description is omitted.

  First, main terms used in the present embodiment will be described.

  The “feature amount” refers to numeric vector data representing the properties of an image.

  “Pre-processing” refers to an image conversion algorithm that is required when performing feature extraction or other pre-processing. For example, edge extraction or color conversion.

  “Control processing” refers to an algorithm for controlling the flow of image recognition processing.

  “Feature extraction” refers to an algorithm for converting an image into a feature value.

  “Learning” refers to an algorithm that performs machine learning of a dictionary to determine what the feature value indicates based on the feature value.

  “Identification” refers to an algorithm for performing identification using learning results. There exists one-to-one in the learning algorithm.

  The “image recognition algorithm” includes the above-described preprocessing, control processing, feature extraction, learning, and identification algorithms.

  The “algorithm component” refers to each specific algorithm belonging to the above-described preprocessing, control processing, feature extraction, learning, and identification. Note that each of the pre-processing, control processing, and feature extraction algorithms is accompanied by a plurality of parameters.

  A “dictionary” refers to data used by machine learning to identify images. One is created for each combination of algorithm, parameter, and recognition target.

  “Learning data” refers to a combination of “image data + category ID (target unit to be identified)” used for machine learning when creating a dictionary.

  “Image data” refers to data obtained by cutting out an object to be identified from an image.

  In this embodiment, after determining a combination of image recognition algorithms using parameters of default values, a dictionary used for image recognition is determined by determining a combination of parameters for the determined algorithm. The number of times of creation is reduced, and the learning data used for creating the dictionary is thinned out, thereby minimizing the time required for evaluating the image recognition performance.

  FIG. 1 is a block diagram showing a schematic configuration of an image recognition algorithm combination selection apparatus according to an embodiment of the present invention. This apparatus is realized using a general-purpose computer (for example, a personal computer (PC) or the like) and software operating on the computer. The computer includes an engineering workstation (EWS) suitable for CAD (Computer Aided Design) and CAE (Computer Aided Engineering). The present embodiment can also be implemented as a program that executes determination of an optimal algorithm candidate, determination of an optimal parameter set, and creation of a dictionary in such a computer.

  As shown in FIG. 1, the image recognition algorithm combination selecting apparatus 100 according to the present embodiment mainly includes an input unit 10, an image recognition algorithm determination unit 20, a learning data list determination unit 30, a dictionary creation unit 40, and a performance evaluation unit 50. , An algorithm component storage unit 60, an algorithm information storage unit 70, and a learning data storage unit 80.

  Learning data and image format information are input to the input unit 10 by the user. The learning data includes image data obtained by cutting out an object to be identified from an image and a category ID representing an identification unit. The image format information includes a color space format and a pixel format of image data. The input learning data and image format information are stored in the learning data storage unit 80.

  The learning data list determination unit 30 extracts a part of image data from a learning data list (described later) in the learning data storage unit 80, and creates a learning data list for creating a dictionary used for evaluation. As will be described in detail later, the learning data list determination unit 30 inputs learning data composed of image data obtained by cutting out an object to be identified from an image and an identification unit, and image format information of the image data. All the combinations of the image format information and the learning data list are used as a learning data list, a part is extracted from the learning data list, and an extracted learning data list for creating a dictionary used for image recognition performance evaluation is created.

  The image recognition algorithm determination unit 20 activates the dictionary creation unit 40 and receives an evaluation result from the performance evaluation unit 50 to determine an optimal algorithm combination and an optimal parameter combination. Although details will be described later, the image recognition algorithm determination unit 20 includes algorithm information including a start condition indicating information for executing the algorithm component and the algorithm component, dependency information indicating a combination of connectable algorithm components, An algorithm candidate list is created using parameter set information representing all parameters used in the algorithm component. Then, an optimal parameter set is determined based on the reevaluation result.

  The dictionary creation unit 40 creates a dictionary for the algorithm candidate list using the default parameter set information and the extracted learning data list. Although the details will be described later, the dictionary creation unit 40 recreates the dictionary corresponding to the optimal algorithm candidate list, and finally recreates the dictionary using all the learning data.

  The performance evaluation unit 50 evaluates the performance of the algorithm candidate list using the parameter set information and the dictionary with respect to the extracted learning data list. Although details will be described later, the performance evaluation unit 50 re-evaluates the performance of the optimal algorithm candidate list using the re-created dictionary.

  The algorithm component storage unit 60 stores a sequence of algorithm information. FIG. 2 is a diagram illustrating an example of algorithm information stored in the algorithm component storage unit 60. As shown in FIG. 2, an algorithm ID, an algorithm component, and a start condition are stored as a set as algorithm information. The sequence of algorithm information is called an algorithm information list. Here, the algorithm component may be the algorithm component (binary) itself or information for acquiring the algorithm component. The start condition is information indicating what algorithm component can be executed when the image format information in the learning data storage unit 80 is. When the algorithm component does not receive image data, “impossible” is entered. In the present embodiment, the color space format and pixel format of input image data are used as image format information. Furthermore, it is assumed that the image formats of all image data are unified.

  In the example shown in FIG. 2, the algorithm ID “0” is “YUV conversion”, the algorithm ID “1” is “edge extraction”, the algorithm ID “10” is “CoHOG”, and the algorithm ID “11” is “Haar-Like”. “AdaBoost” is stored as the algorithm ID “21”, “SVM” is stored as the algorithm ID “22”, and the like. According to the example illustrated in FIG. 2, the algorithm component “YUV conversion” indicates that it can be executed when the image format information is “RGB444”. Further, the algorithm component “edge extraction” indicates that it can be executed when the image format information is “Y”.

  The algorithm information storage unit 70 stores dependency relationship information and parameter set information. Here, the dependency relationship is a combination of connectable algorithms and does not include a relationship that forms a loop. FIG. 3 is a diagram illustrating an example of dependency relationship information. In the example shown in FIG. 3, when the algorithm ID “0” is the preceding algorithm part and the algorithm ID “10” is the latter algorithm part, the connection is possible. Similarly, connection is possible even when the algorithm ID “1” is the preceding algorithm part and the algorithm ID “10” is the latter algorithm part.

  Further, the algorithm information storage unit 70 stores information on parameters that can be used for each algorithm. If a combination of parameters used in one algorithm is a parameter set, each algorithm has a plurality of parameter sets.

  FIG. 4 is a diagram illustrating an example of parameter information. As shown in FIG. 4, the parameter information stores an algorithm ID, a parameter set ID, and a parameter.

  In the example shown in FIG. 4, two parameter sets that can be used by the algorithm ID “10” are stored as the parameter set ID “1” and the parameter set ID “2”. Further, two parameter sets that can be used by the algorithm ID “23” are stored as a parameter set ID “1” and a parameter set ID “2”. In this embodiment, the parameter set ID “1” is used as a default value.

  Specific parameters of the parameter set ID “1” of the algorithm ID “10” include “direction quantization number = 4”, “region division number (x) = 1”, and “region division number (y) = 1”. As specific parameters of the parameter set ID “2” stored in the algorithm ID “10”, “direction quantization number = 4”, “region division number (x) = 2”, “region division number (y) = 3 "is stored.

  Similarly, as specific parameters of the parameter set ID “1” of the algorithm ID “23”, “basic number = 100”, “positive class number = 20”, “negative class number = 20”, “similar range value = 0.2 ”is stored, and specific parameters of the parameter set ID“ 2 ”of the algorithm ID“ 23 ”are“ basic number = 200 ”,“ positive class number = 20 ”,“ negative class number = 20 ”, “Similar range value = 0.3” is stored.

  The learning data storage unit 80 stores learning data and image format information prepared by the user. FIG. 5 is a diagram illustrating an example of learning data information. As shown in FIG. 5, learning data ID, category ID, image data, and image format information are stored as learning data information, all of which are called a learning data list. Here, the learning data ID is an ID assigned to a combination of a category ID and image data. The category ID refers to data used as a correct answer of an identification result for evaluating machine learning or a dictionary for each target to be identified. The image data is obtained by cutting out an object to be identified from an image, and may be binary data itself or information for acquiring binary data, for example.

  The image format information is the color space format and the pixel format of the input image data. In this embodiment, the image formats of all the image data when performing one combination optimization process are unified. That is, the combination optimization processing of a certain algorithm targets all image data in the “RGB444” format, and the combination optimization processing of another algorithm targets all image data in the “Y” format. In the combination optimization process, all image data in the “YUV” format is targeted.

  In the example shown in FIG. 5, the category ID “1” and the image data “person image 1” are stored as the learning data ID “1”, and the category ID is set as the learning data ID “3107”. “2” and image data “background image 1” are stored. Furthermore, “RGB444” is stored as image format information.

<Algorithm candidate list>
When creating and evaluating a dictionary for an image recognition algorithm, an algorithm component and a parameter set are used in combination. This combination is called an “algorithm candidate”, and a sequence of algorithm parts and parameter sets is called an “algorithm candidate list”. The algorithm candidates and the algorithm candidate list are created by the image recognition algorithm determination unit 20. FIG. 6 is a diagram illustrating an example of the algorithm candidate list. In the example shown in FIG. 6, a combination of algorithm ID “7” and parameter set ID “1” is an algorithm candidate. The algorithm candidate list is composed of four columns including this set.

<Algorithm performance information list>
It is necessary to evaluate whether or not the algorithm combination configuration is optimal under a certain index. In this embodiment, the evaluation result is performed by numerical data such as a correct answer rate (number of data successfully identified / number of learning data used for evaluation). In the present embodiment, the combination of the algorithm candidate list and the evaluation result is referred to as algorithm performance information. The algorithm performance information is created by the image recognition algorithm determination unit 20. FIG. 7 is a diagram illustrating an example of the algorithm performance information list. In the example shown in FIG. 7, for an algorithm candidate list composed of algorithm IDs “0”, “7”, “10”, “23” and parameter set ID “1”, an evaluation result of 1898/2000 is used as algorithm performance information. Are listed.

  A flow of processing in the image recognition algorithm combination selecting apparatus 100 configured as described above will be described. In this embodiment, the number of times of dictionary creation is divided into two steps of “determination of algorithm combination” and “determination of parameter combination” so that the number of dictionary creation times is reduced from the order of (m * n) times to (m + n). ) Reduce to order times. Here, m represents the number of algorithm combinations, and n represents the number of parameter combinations. The combination of algorithms is the most important whether an image can be identified. Accordingly, algorithm combinations are determined using parameters of default values, and parameters are determined for the determined algorithm.

  In the following, processing will be broadly described until algorithm combination determination, parameter combination determination, algorithm candidate list and dictionary output.

<Processing until determination of algorithm combination>
FIG. 8 is a flowchart showing the flow of processing up to the determination of the algorithm combination.

  First, the user inputs all learning data (step S801).

  Next, the user inputs an image format (step S802).

  Next, from the algorithm information list, a startable algorithm information list in which all algorithm information having the same start condition of the algorithm information and the input image format information is extracted is created (step S803). For example, based on the algorithm information list stored in the algorithm component storage unit 60 shown in FIG. 2 and the image format information stored in the learning data storage unit 80 shown in FIG. 5, the startable algorithm information list shown in FIG. Is created.

  Next, it is determined whether or not the startable algorithm information list is empty (step S804).

  If the startable algorithm information list is empty (Yes in step S804), the process in the image recognition algorithm combination selection device is terminated. If the startable algorithm information list is not empty (No in step S804), empty algorithm performance is obtained. An information list (see FIG. 7) is created (step S805).

  Next, one piece of algorithm information is extracted from the startable algorithm information list (see FIG. 9) and deleted from the original list (step S806).

  Next, an empty algorithm candidate is created, and the algorithm ID of the created algorithm candidate (see FIG. 6) is set as the algorithm ID of the algorithm information extracted in step S806. Further, the parameter set ID of the algorithm candidate is set to 1 (default value) (step S807). From the example shown in FIG. 9, algorithm candidates with algorithm ID “0” and parameter set ID “1” are created.

  Next, an empty algorithm candidate list is created, and the algorithm candidate created in step S807 is added to the bottom of the created algorithm candidate list (step S808). In the example of step S807, only algorithm candidates consisting of the algorithm ID “0” and the parameter set ID “1” are described in the algorithm candidate list.

  Next, process 1 is executed using the algorithm candidate list as an argument (step S809). In step S809, a dictionary is created for the m algorithm candidate lists. Here, the process 1 is to create a new algorithm candidate list by sequentially adding connectable algorithm candidates to the given algorithm candidate list under the conditions given by the dependency relationship information. .

  Below, an example of the detail of the process 1 is described.

Process 1 (argument: algorithm candidate list) {
f1.1. Get the algorithm candidate at the bottom of the algorithm candidate list.

    f1.2. Get the algorithm ID from the obtained algorithm candidates.

    f1.3. Get all dependency information where the algorithm ID obtained from the dependency information list is the same as the previous algorithm ID.

    f1.4. Extract the latter algorithm ID from all the acquired dependency information and create the latter algorithm ID list.

    f1.5. If the list is not empty:

    f1.6. Execute the following for each element of the latter algorithm ID.

    f1.7. Create an empty algorithm candidate.

    f1.8. Set the latter algorithm ID to the algorithm ID of the candidate algorithm created.

    f1.9. Furthermore, the parameter set ID of the algorithm candidate is set to 1 (default value).

    f1.10. Make a copy of the argument algorithm candidate list.

    f1.11. Add the created algorithm candidates to the bottom of the algorithm candidate list created as a copy.

    f1.12. Process 1 is executed with the algorithm candidate list created as a copy as an argument.

f1.13. If the list is empty
f1.14. Process 2 (described later) is executed with the argument algorithm candidate list as an argument.
}

  FIG. 10 is a diagram illustrating an example of data generated every time the process 1 is recursively called. FIG. 11 is a diagram for explaining the work in the process 1, and is a diagram for explaining the calculation of the lower data within the dotted line in FIG.

  In process 2, when there are no more algorithm candidates that can be connected in process 1, dictionary creation and performance evaluation are performed based on the algorithm candidate list at that time. Below, an example of the detail of the process 2 is described.

Process 2 (argument: algorithm candidate list) {
f2.1. Create a learning data list by extracting some learning data for each category ID from the learning data list. For example, 40% of some learning data is randomly extracted from learning data having the same category ID for each category ID. It is necessary that the number of image data included in the category does not become zero. FIG. 12 is a diagram illustrating an example of the learning data list after extraction.

    f2.2. For the algorithm candidate list, create a dictionary using the created learning data list, evaluate the performance, and calculate the evaluation result. For example, a general method such as cross-validation can be used for evaluating the algorithm candidate list. FIG. 13 is a diagram illustrating an example of calculation of an evaluation result. The evaluation result can include a numerical value capable of quantitatively comparing the identification performance such as the accuracy rate (number of data successfully identified / number of learning data used for evaluation).

    f2.3. Create an empty algorithm performance information list, and set the argument algorithm candidate list and the calculated evaluation result as setting values.

f2.4. Add the created algorithm performance information to the created algorithm performance information list.
}

  Next, it is determined whether or not the startable algorithm information list is empty (step S810).

  If the startable algorithm information list is empty (Yes in step S810), the algorithm candidate list having the best evaluation result is determined as the optimum algorithm candidate list from the algorithm performance information list created in step S806 (step S811). The process up to the determination of the algorithm combination is completed. FIG. 14 is a diagram illustrating an example of an optimal algorithm candidate list obtained by the processing up to step S811.

  On the other hand, if the startable algorithm information list is not empty (No in step S810), the process proceeds to step S806.

<Processing until determination of parameter combination>
FIG. 15 is a flowchart showing the flow of processing up to the determination of the parameter combination.

  First, an empty algorithm performance information list is created (step S1501).

  Next, a copy of the optimal algorithm candidate determined in step S811 is created (step S1502).

  Next, the parameter set of each algorithm candidate in the algorithm candidate list created as a copy is emptied (step S1503).

  Next, process 3 is executed using the algorithm candidate list created as a copy as an argument (step S1504). In step S1504, a dictionary is created for n parameter combinations. Process 3 is to sequentially set the parameters given by the parameter information for the given algorithm candidate list corresponding to each algorithm candidate list. Below, an example of the detail of the process 3 is described.

Process 3 (argument: algorithm candidate list) {
f3.1. Get one algorithm candidate whose parameter set ID is empty from the algorithm candidate list.

    f3.2. Get the algorithm ID from the obtained algorithm candidates.

    f3.3. If there is an algorithm candidate included in the argument algorithm candidate list with an empty parameter set ID.

    f3.4. From the parameter information, obtain all parameter sets with the same algorithm ID as the obtained algorithm ID and create a parameter set list.

    f3.5. Execute the following for each element of the parameter set list.

    f3.6. Get parameter set ID from parameter set.

    f3.7. Make a copy of the argument algorithm candidate list.

    f3.8. From the algorithm candidate list created as a copy, obtain an algorithm candidate with the same algorithm ID as the obtained algorithm ID.

    f3.9. Set the parameter set ID obtained as the parameter set ID value of the algorithm candidate obtained in f3.8.

    f3.10. Execute process 3 with the algorithm candidate list created as a copy as an argument.

    f3.11. If none of the algorithm candidates included in the argument algorithm candidate list has an empty parameter set ID.

f3.12. Process 4 is executed with the argument algorithm candidate list as an argument.
}

  In the process 4, when there are no more parameters that can be set in the process 3, the dictionary is created and the performance is evaluated based on the algorithm candidate list at that time. Below, an example of the detail of the process 4 is described.

Process 4 (argument: algorithm candidate list) {
f4.1. Create a learning data list by extracting some learning data for each category ID from the learning data list. Some of the learning data is the same as described above.

    f4.2. For the algorithm candidate list, create a dictionary using the created learning data list, evaluate the performance, and calculate the evaluation result.

    f4.3. Create empty algorithm performance information and set the argument algorithm candidate list and the calculated evaluation result as the setting values.

f4.4. Add the created algorithm performance information to the created algorithm performance information list.
}

  FIG. 16 is a diagram illustrating an example of data generated every time the process 3 is recursively called. FIG. 17 is a diagram for explaining the work in the process 3.

  Next, the algorithm candidate list with the best evaluation result is determined as the optimum algorithm candidate list from the algorithm performance information list created in step S1501 (step S1505), and the process up to the determination of the parameter combination is completed. FIG. 18 is a diagram illustrating an example of the optimum algorithm candidate list obtained by the processing up to step S1505.

<Processing up to algorithm candidate list and dictionary output>
FIG. 19 is a flowchart showing the flow of processing up to the algorithm candidate list and dictionary output.

  First, a dictionary is created using the learning data list for the optimal algorithm candidate list determined in step S1505 (step S1901).

  Next, the optimal algorithm candidate list determined in step S1505 and the dictionary created in step S1901 are output (step S1902). FIG. 20 is a diagram illustrating a final dictionary created using all learning data. As described above, the combination of the determined algorithms and the parameters used in each algorithm are optimized, and a dictionary is created, so that an image recognizer having an optimal combination can be configured.

<Modification>
The following modifications can also be applied to the present embodiment.

(1) Although an optimal algorithm candidate list is determined in step S811 or step S1505, a list may be displayed.
(2) In step S809 and step S1504, the algorithm is realized by recursive processing. However, this portion can be converted to repetitive processing by using well-known intermediate data of a tree structure or the like.
(3) In step S1901, the dictionary is created using the entire learning data list. However, the dictionary may be created using a sufficient amount of the learning data list instead of the entire learning data list.

  As described above, according to the present embodiment, the time required to mechanically determine the combination configuration of the image recognition algorithms is “algorithm combination (m ways)” * “parameter combination (n ways)”. The order of (m * n) is reduced to the order of (m + n) of “algorithm combination (m ways)” + “parameter combination (n ways)”. A combination of algorithms for recognition, a combination of parameters for each algorithm, and a dictionary can be determined.

  Although several embodiments of the present invention have been described, these embodiments are presented by way of example and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 100 ... Image recognition algorithm combination selection apparatus 10 ... Input part 20 ... Image recognition algorithm determination part 30 ... Learning data list determination part 40 ... Dictionary creation part 50 ... Performance evaluation part 60- ..Algorithm component storage unit 70 ... Algorithm information storage unit 80 ... Learning data storage unit

Claims (9)

An apparatus for selecting a combination of algorithm parts for image recognition,
Input learning data consisting of image data obtained by cutting out an object to be identified from an image and an identification unit, and image format information of the image data, and learn all sets of the learning data, the image data, and the image format information A learning data list determination unit for creating a data list, extracting a part from the learning data list, and creating an extracted learning data list for creating a dictionary used for performance evaluation of image recognition;
Algorithm information including a start condition that represents the algorithm component and information for enabling execution of the algorithm component, dependency information that represents a combination of the algorithm components that can be connected, and all parameters used in the algorithm component An image recognition algorithm determination unit that creates an algorithm candidate list using parameter set information to represent,
A dictionary creation unit that creates a dictionary using the parameter set information of default values and the extracted learning data list for the algorithm candidate list,
A performance evaluation unit that evaluates the algorithm candidate list using the parameter set information and the dictionary for the extracted learning data list,
The image recognition algorithm determination unit sets the parameters of the parameter set information, with the algorithm candidate list having the best evaluation result of the performance evaluation unit as the optimal algorithm candidate list,
The dictionary creation unit recreates the dictionary corresponding to the optimal algorithm candidate list,
The performance evaluation unit re-evaluates the performance of the optimal algorithm candidate list using the re-created dictionary,
The image recognition algorithm determination unit determines an optimal parameter set based on the reevaluation result,
The dictionary creation unit is an image recognition algorithm combination selection device that finally recreates the dictionary using all the learning data.   The learning data is extracted at random from the learning data having the same category ID so that the number of image data included in the category ID does not become zero for each category ID representing the identification unit. The image recognition algorithm combination selection apparatus according to claim 1, wherein the ratio is extracted. An input unit for inputting the learning data and the image format information;
An algorithm component storage section for storing a column of the algorithm information;
An algorithm information storage unit for storing the dependency relationship information and the parameter set information;
The image recognition algorithm combination selection apparatus according to claim 1, further comprising a learning data storage unit that stores the learning data list and the image format information.   The image recognition algorithm combination selection device according to any one of claims 1 to 3, wherein one dictionary is created for a combination of an algorithm, a parameter, and a recognition target.   The image recognition algorithm combination selection according to any one of claims 1 to 4, wherein the dependency relationship information does not include a combination in which the connection between the algorithm component in the previous stage and the algorithm component in the subsequent stage is a loop. apparatus.   The image according to any one of claims 1 to 5, wherein the image format information includes a color space format and a pixel format of the image data, and the image format information of all the image data is unified. Recognition algorithm combination selection device.   In determining the optimum algorithm candidate list, an algorithm performance information list including a combination of the algorithm candidate list and the evaluation result is created, and the algorithm candidate list having the best evaluation result is used as the optimum algorithm candidate list. The image recognition algorithm combination selection apparatus according to any one of claims 1 to 6.   The image recognition algorithm combination selection device according to any one of claims 1 to 7, wherein the algorithm component is binary data or information for acquiring the algorithm component.   The image recognition algorithm combination selection according to any one of claims 1 to 8, wherein the evaluation result is numerical data with a correct answer rate expressed by (number of data successfully identified / number of learning data used for evaluation). apparatus.

Images