Methods, apparatus, equipment and media for regression analysis of convolutional neural network behavior.

Abstract

The present invention is mainly used in the field of image processing technology. The present invention discloses a method, device, equipment and medium for analyzing the behavior of a regression analysis convolutional neural network. The method includes: first, inputting an input image into a regression analysis convolutional neural network to obtain a desired output. Then, creating a population matrix in a randomly initialized manner, and each individual is a binary filter. Through preset iteration and screening conditions, the population matrix is ​​updated to determine the optimal filter. Then, the input image is processed with the optimal filter to obtain a target image, and the target image is input again into the regression analysis convolutional neural network to obtain a target output. Finally, by comparing the difference between the target image and the original image, and the difference between the target output and the expected output, the predicted behavior of the regression analysis convolutional neural network is analyzed. By identifying the relationship between the input value and the output value of the regression analysis convolutional neural network, it helps users understand the prediction basis of the neural network model.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method, apparatus, device, and medium for regressive analysis of convolutional neural network behavior. Background Art

[0002] In related technical fields, due to the black-box nature of deep convolutional neural networks, it is usually difficult to understand why a deep convolutional network outputs a specific predicted value when given a specific image as input. This limits the application of deep convolutional neural networks in some high-risk tasks. Moreover, the interpretability of deep convolutional neural networks is usually only applicable to deep convolutional neural networks in classification tasks, but not to deep convolutional neural networks in regression analysis tasks. Summary of the Invention

[0003] This invention provides a method, apparatus, device, and medium for analyzing the behavior of a convolutional neural network in regression analysis. By identifying the relationship between the input and output values ​​of the convolutional neural network in regression analysis, it helps users understand the predictive basis of the neural network model.

[0004] This invention provides a method for regressive analysis of the behavior of convolutional neural networks, the method comprising:

[0005] The input image and the regression analysis convolutional neural network to be parsed are obtained, and the input image is input into the regression analysis convolutional neural network for processing to obtain the desired output value;

[0006] A population matrix is ​​created using random initialization, where each element in the population matrix is ​​a binary filter for behavioral analysis of the regression analysis convolutional neural network.

[0007] The population matrix is ​​updated based on a preset number of iterations, and the optimal element individual is determined from the updated population matrix according to preset screening conditions.

[0008] The binary filter corresponding to the optimal element is convolved with the input image to obtain the target image, and the target image is then input into the regression analysis convolutional neural network for processing to obtain the target output value.

[0009] The relationship between the input and output values ​​of the regression analysis convolutional neural network is determined based on the first difference and the second difference to analyze the prediction behavior of the regression analysis convolutional neural network. The first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

[0010] Furthermore, the step of updating the population matrix based on a preset number of iterations and determining the optimal individual from the updated population matrix according to preset screening conditions includes:

[0011] Update the number of individuals and the characteristics of individuals in the current population matrix;

[0012] Determine the fitness of each individual element in the updated current population matrix;

[0013] Based on the fitness of each individual element, multiple selected individual elements are chosen from the current population matrix;

[0014] A new population matrix is ​​generated based on the multiple selected element individuals;

[0015] The new population matrix is ​​used as the current population matrix, and the process returns to the step of updating the number of individual elements and the characteristics of individual elements in the current population matrix.

[0016] When the number of times the new population matrix is ​​generated is equal to the preset number of iterations, the new population matrix is ​​used as the updated population matrix, and the element with the best fitness in the updated population matrix is ​​used as the optimal element.

[0017] Furthermore, determining the fitness of each individual element in the updated current population matrix includes:

[0018] The input image is convolved with each binary filter in the population matrix to obtain images after multiple convolution operations;

[0019] The image after each convolution operation is input into the regression analysis convolutional neural network for processing to obtain the actual output value corresponding to each image after the convolution operation;

[0020] The expected output value is compared with each of the actual output values ​​to determine the absolute error value corresponding to each of the binary filters, wherein the absolute error value is used as the fitness of the individual element.

[0021] Furthermore, the step of selecting multiple selected element individuals from the current population matrix based on the fitness of each element individual includes:

[0022] Obtain the first candidate element individual and the second candidate element individual from the current population matrix;

[0023] Determine whether the fitness of the first candidate element is greater than that of the second candidate element. If yes, then the second candidate element is selected as the filtered element. If no, then the first candidate element is selected as the filtered element.

[0024] Return to the step of obtaining the first candidate element individual and the second candidate element individual from the current population matrix, until the number of the filtered element individuals is equal to the preset number threshold.

[0025] Furthermore, updating the number of individuals and the characteristics of individuals in the current population matrix includes:

[0026] The current population matrix is ​​divided into multiple element groups, and each element group includes a first element individual and a second element individual;

[0027] Based on a first preset probability value, a target element group is determined from the plurality of element groups, and a plurality of third element individuals are generated according to the characteristics of the first element individual and the characteristics of the second element individual in each target element group.

[0028] Based on the second preset probability value, a mutation operation is performed on each of the third element individuals to obtain a new element individual;

[0029] The updated current population matrix is ​​generated based on the new element individuals.

[0030] Furthermore, the step of determining a target element group from the plurality of element groups based on a first preset probability value, and generating a plurality of third element individuals according to the characteristics of the first element individuals and the characteristics of the second element individuals within each target element group, includes:

[0031] Based on a first preset probability value, a target element group is determined from the plurality of element groups;

[0032] The first and second individual elements in the target element group are respectively converted into a first vector and a second vector, and both the first vector and the second vector include multiple parameters.

[0033] At least one parameter in the first vector is swapped with the target parameter in the second vector, and the first and second vectors after the parameter swap are respectively converted into the third element individual.

[0034] Furthermore, the step of performing a mutation operation on each of the third element individuals based on a second preset probability value to obtain a new element individual includes:

[0035] The third element individual is converted into a third vector, which includes multiple parameters;

[0036] Based on the second preset probability value, the value of the target parameter among the multiple parameters is adjusted to obtain the third vector after parameter adjustment and convert it into the new element individual.

[0037] The present invention also provides a regression analysis convolutional neural network behavior parsing device, the device comprising:

[0038] The first processing module is used to acquire the input image and the regression analysis convolutional neural network to be parsed, and to input the input image into the regression analysis convolutional neural network for processing to obtain the desired output value;

[0039] A generation module is used to create a population matrix in a random initialization manner, wherein each element in the population matrix is ​​a binary filter for the behavior analysis of the regression analysis convolutional neural network.

[0040] An iteration module is used to update the population matrix based on a preset number of iterations and determine the optimal element individual from the updated population matrix according to preset screening conditions.

[0041] The second processing module is used to perform a convolution operation between the binary filter corresponding to the optimal element individual and the input image to obtain the target image, and input the target image into the regression analysis convolutional neural network for processing to obtain the target output value;

[0042] The parsing module is used to determine the relationship between the input and output values ​​of the regression analysis convolutional neural network based on a first difference and a second difference, so as to analyze the prediction behavior of the regression analysis convolutional neural network, wherein the first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

[0043] The present invention also provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the regression analysis convolutional neural network behavior parsing method as described in any of the preceding claims.

[0044] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the regression analysis convolutional neural network behavior parsing method as described in any of the preceding claims.

[0045] The present invention has at least the following beneficial effects:

[0046] This technical solution analyzes the impact of binary filters on the predictive output of a regression analysis convolutional neural network. Specifically, a first difference reveals the input image feature changes that lead to prediction variations, while a second difference reveals how these input image feature changes affect the output of the regression analysis convolutional neural network. Analyzing both the first and second differences helps reveal the working mechanism of the regression analysis convolutional neural network in regression analysis tasks, thereby improving the model's interpretability. Furthermore, by transforming the complex decision-making process of the regression analysis convolutional neural network into observable image feature changes, users can more intuitively understand how the model makes specific predictions, thus helping them comprehend the predictive behavior of the regression analysis convolutional neural network. Attached Figure Description

[0047] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0048] Figure 1 This is a flowchart illustrating the steps of a regression analysis method for interpreting the behavior of convolutional neural networks.

[0049] Figure 2 This is a flowchart of step S103 in a regression analysis method for analyzing the behavior of convolutional neural networks.

[0050] Figure 3 This is a flowchart of step S201 in a regression analysis method for interpreting the behavior of convolutional neural networks.

[0051] Figure 4 This is a flowchart of step S202 in a regression analysis method for parsing the behavior of convolutional neural networks.

[0052] Figure 5 This is a flowchart of step S203 in a regression analysis method for interpreting the behavior of convolutional neural networks.

[0053] Figure 6 This is a schematic diagram of a device for analyzing the behavior of a convolutional neural network in regression analysis.

[0054] Figure 7 This is a schematic diagram of the structure of an electronic device. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] It should be noted that deep convolutional neural networks (DCNNs) have been widely used in image-based regression analysis tasks in related technical fields. However, due to the black-box nature of DCNNs, given a specific image as input, it is often difficult to understand why the DCNN outputs a specific predicted value, thus limiting the application of DCNNs in some high-risk tasks.

[0057] To address the interpretability of deep convolutional neural networks (DCNNs), several techniques exist for understanding their behavior, including category activation map-based techniques and perturbation-based techniques. However, these techniques are primarily aimed at DCNNs in classification tasks, while techniques for interpreting DCNNs in regression analysis tasks are relatively scarce. There is a fundamental difference between DCNNs used in classification and regression analysis tasks: the output of a DCNN in classification is a discrete category label, while the output of a DCNN in regression analysis is a continuous numerical value.

[0058] To help users understand the predictive behavior of deep convolutional neural networks (DCNNs) in regression analysis tasks, this application provides a method, apparatus, device, and medium for analyzing the behavior of DCNNs in regression analysis. This solution utilizes a genetic algorithm to evolve the interpretability filters of the DCNN, so that when the result of convolution with an image using the evolved filters is used as input to the DCNN, the DCNN can output a specific value. This identifies patterns in the image that cause the DCNN to output a specific predicted value, helping users understand why the DCNN makes a specific prediction given a specific input. The technical solution for analyzing the behavior of DCNNs in regression analysis includes the following embodiments.

[0059] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a regression analysis method for interpreting the behavior of convolutional neural networks.

[0060] This embodiment provides a regression analysis method for parsing the behavior of convolutional neural networks, including:

[0061] S101. Obtain the input image and the regression analysis convolutional neural network to be parsed, and input the input image into the regression analysis convolutional neural network for processing to obtain the desired output value.

[0062] S102. Create a population matrix using random initialization. Each element in the population matrix is ​​a binary filter for analyzing the behavior of a convolutional neural network in regression analysis.

[0063] S103. Update the population matrix based on the preset number of iterations, and determine the optimal element individual from the updated population matrix according to the preset screening conditions.

[0064] S104. Perform a convolution operation between the binary filter corresponding to the optimal element individual and the input image to obtain the target image, and then input the target image into the regression analysis convolutional neural network for processing to obtain the target output value.

[0065] S105. Determine the relationship between the input and output values ​​of the regression analysis convolutional neural network based on the first difference and the second difference, so as to analyze the prediction behavior of the regression analysis convolutional neural network. The first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

[0066] In some embodiments, the binary filter for regression analysis of convolutional neural network behavior analysis includes... A binary filter corresponding to a matrix of values.

[0067] In step S101 of some embodiments, an automated script can be used to retrieve the input image and the corresponding DCNN model from the database, or the input image can be uploaded through a user interface, and a pre-trained DCNN model can be selected for analysis.

[0068] In step S102 of some embodiments, an initial population matrix is ​​randomly generated, wherein the filter parameters of each individual are random binary values, and optionally, filters that have performed well in the previous optimization process are used as part of the initial population.

[0069] In step S103 of some embodiments, the population is updated using standard operations of genetic algorithms (selection, crossover, mutation). Optionally, fitness ratio selection is introduced, prioritizing individuals with high fitness for crossover and mutation. Furthermore, the probabilities of crossover and mutation are dynamically adjusted during the iteration process to explore and develop different search regions.

[0070] In step S105 of some embodiments, the specific ways of parsing the prediction behavior of the regression analysis convolutional neural network include: calculating statistical data (such as mean, standard deviation) of the difference to analyze the prediction behavior, using machine learning methods to model the difference to predict the behavior of DCNN under different inputs, or displaying the difference through visualization tools to help users intuitively understand the prediction behavior of DCNN.

[0071] Please refer to Figure 2 , Figure 2 This is a flowchart of step S103 in a regression analysis method for interpreting the behavior of convolutional neural networks.

[0072] In some embodiments, step S103 includes:

[0073] S201. Update the number of individual elements and the characteristics of individual elements in the current population matrix.

[0074] S202. Determine the fitness of each individual element in the updated current population matrix.

[0075] S203. Based on the fitness of each individual element, select multiple selected individuals from the current population matrix.

[0076] S204. Generate a new population matrix based on multiple screened individual elements.

[0077] S205. Use the new population matrix as the current population matrix and return to step S201.

[0078] S206. When the number of times a new population matrix is ​​generated is equal to the preset number of iterations, the new population matrix is ​​used as the updated population matrix, and the element with the best fitness in the updated population matrix is ​​used as the best element.

[0079] In one specific embodiment, let t = 1, 2, 3, ..., T, where T is the maximum set number of iterations. First, an initial population matrix is ​​generated, containing multiple individual elements. Each individual element represents a binary filter. Optionally, a wider feature space can be explored by increasing the population size or adjusting the size and shape of each filter.

[0080] Next, each binary filter is applied to the input image, and the resulting image is processed using a DCNN. Then, the error between the image processed by each filter and the desired output is calculated. This error will be used as the fitness evaluation for each individual.

[0081] Then, based on fitness evaluation, the top 10% of individuals with the smallest errors are selected. These individuals are considered the optimal binary filter. The selected individuals are then subjected to crossover (combining the features of two excellent individuals to generate new individuals) and mutation (randomly changing the features of some individuals) operations to generate a new population matrix. This process simulates natural selection and genetic mutation in the hope of producing a better binary filter. The newly generated population matrix replaces the old population matrix, and the above process is repeated. This process continues until a preset number of iterations is reached. When the number of times the new population matrix is ​​generated equals the preset number of iterations, the new population matrix is ​​used as the updated population matrix, and the individual with the best fitness in the updated population matrix is ​​taken as the optimal individual.

[0082] Finally, after all iterations are completed, the individual with the highest fitness is selected, which is the binary filter with the smallest error.

[0083] It is understood that this embodiment optimizes the prediction performance of the deep convolutional neural network by iteratively updating and filtering individuals in the population matrix, and finally finds the filter with the best parameter configuration to improve the accuracy and interpretability of the model.

[0084] Please refer to Figure 3 , Figure 3 This is a flowchart of step S201 in a regression analysis method for interpreting the behavior of convolutional neural networks.

[0085] In some embodiments, step S201 includes:

[0086] S301. Divide the current population matrix into multiple element groups, each element group including the first element individual and the second element individual.

[0087] S302. Based on the first preset probability value, determine the target element group from multiple element groups, and generate multiple third element individuals according to the characteristics of the first element individual and the characteristics of the second element individual in each target element group.

[0088] S303. Based on the second preset probability value, perform a mutation operation on each third element individual to obtain a new element individual.

[0089] S304. Generate an updated current population matrix based on the new element individuals.

[0090] Specifically, the specific implementation of step S302 includes: converting the population matrix... Divided in order Yes, recorded as ,in This indicates a floor operation. For each element group... Crossover operations are performed with a first preset probability pc.

[0091] In some embodiments, the specific implementation of step S302 is as follows:

[0092] Based on a first preset probability value, a target element group is determined from multiple element groups; the first and second element individuals in the target element group are converted into a first vector and a second vector, respectively, and both the first and second vectors include multiple parameters; at least one parameter in the first vector is swapped with the target parameter in the second vector, and the first and second vectors after parameter swapping are converted into a third element individual, respectively.

[0093] For example, reshape both individual elements in each element group (i.e., the first element and the second element) into one. The vector is then used to randomly select a position from the vector, and the parts after that position of two individual elements are swapped to obtain a third individual element. These two third individual elements are then reconstructed back into their original positions. Filter.

[0094] In some embodiments, the specific implementation of step S302 is as follows:

[0095] The third element individual is converted into a third vector, which includes multiple parameters; based on the second preset probability value, the value of the target parameter among the multiple parameters is adjusted to obtain the parameter-adjusted third vector and convert it into a new element individual.

[0096] For example, first randomly select L parameters from the third vector (L≤ Then, a mutation operation is performed on each selected parameter with a second preset probability pm. That is, if the original value of a parameter is 0, it will become 1 with probability pm after the mutation operation, and if the original value of a parameter is 1, it will become 0 with probability pm after the mutation operation.

[0097] Understandably, this embodiment optimizes the binary filter by simulating a genetic evolution process. First, the population matrix is ​​divided into multiple element groups, each containing two individuals. Then, a crossover operation is performed with a preset probability to generate new individuals with a third element. Next, these new individuals undergo mutation to increase population diversity. By iteratively updating the population matrix, the individual with the highest fitness is finally selected. This optimization method improves the filter's performance, enhances the model's ability to recognize image features, and thus improves the prediction accuracy and interpretability of deep convolutional neural networks in regression analysis tasks. Furthermore, through an automated evolutionary strategy, this approach helps reduce manual intervention and improves the efficiency of model training.

[0098] Please refer to Figure 4 , Figure 4 This is a flowchart of step S202 in a regression analysis method for analyzing the behavior of convolutional neural networks.

[0099] In some embodiments, step S202 includes:

[0100] S401. Perform convolution operations on the input image and each binary filter in the population matrix to obtain an image after multiple convolution operations.

[0101] S402. Input the image after each convolution operation into the regression analysis convolutional neural network for processing to obtain the actual output value corresponding to the image after each convolution operation.

[0102] S403. Compare the expected output value with each actual output value to determine the absolute error value corresponding to each binary filter, wherein the absolute error value is used as the fitness of the individual element.

[0103] In one specific embodiment, a population containing N individuals is first randomly initialized, denoted as . ,in Let f represent the nth individual in the 0th generation population. Each individual is a binary filter with a value of 0 or 1. The fitness calculation process for each individual in the population is as follows: First, convolve the individual with the input image, denoted as x∗f. Then, use the convolved image as the input to a deep convolutional neural network F(). Next, the output F(x∗f) is obtained through the deep convolutional neural network F(). Finally, calculate the absolute error between the expected output value y and F(x∗f): |y−F(x∗f)|. This absolute error is the fitness of individual f.

[0104] Understandably, this embodiment calculates the absolute error fitness of each filter by convolving the input image with each binary filter and then feeding it into a convolutional neural network to obtain the actual output value. This output value is then compared with the expected output value. This process achieves a quantitative evaluation of filter performance, optimizes the filters to improve the network's accuracy in recognizing image features, and thus enhances the model's prediction accuracy in regression analysis tasks.

[0105] Please refer to Figure 5 , Figure 5 This is a flowchart of step S203 in a regression analysis method for interpreting the behavior of convolutional neural networks.

[0106] In some embodiments, step S203 includes:

[0107] S501. Obtain the first and second candidate elements from the current population matrix.

[0108] S502. Determine whether the fitness of the first candidate element individual is greater than the fitness of the second candidate element individual. If yes, then the second candidate element individual is selected as the element individual after screening. If no, then the first candidate element individual is selected as the element individual after screening.

[0109] S503. Return to step S501 until the number of filtered individual elements equals the preset quantity threshold.

[0110] In one specific embodiment, firstly, the current population matrix of the (t-1)th generation population is... Two individuals are randomly selected from the data. and These are respectively designated as the first candidate element individual and the second candidate element individual. Then, the first candidate element individual... Second candidate element individual Compare fitness levels, if Fitness less than or equal to The fitness, then Selected for the t-th generation population, otherwise The selected individuals are chosen. This selection process is repeated N times to ensure that the number of individuals in the t-th generation population is N. Finally, the current population matrix of the (t-1)-th generation population is... Select N individuals to obtain the new population matrix of the t-th generation. .

[0111] Understandably, this embodiment compares the fitness of two candidate individuals and selects the individual with lower fitness as the filtered element, thus achieving survival of the fittest in the population. This process is repeated until a preset screening threshold is reached, ensuring that the population retains only the individuals with the highest fitness. This method improves the overall quality of the population and promotes the improvement of model performance, especially in optimization problems, where it helps to quickly converge to a better solution.

[0112] It is understood that any of the above embodiments can analyze the impact of binary filters on the prediction output of the regression analysis convolutional neural network. Specifically, the first difference reveals the changes in input image features that cause the prediction change, and the second difference reveals how these changes in input image features affect the output of the regression analysis convolutional neural network. Analyzing the first and second differences helps to reveal the working mechanism of the regression analysis convolutional neural network in regression analysis tasks, thereby improving the interpretability of the model. Furthermore, by transforming the complex decision-making process of the regression analysis convolutional neural network into observable changes in image features, users can more intuitively understand how the model makes specific predictions, thus helping users understand the predictive behavior of the regression analysis convolutional neural network.

[0113] Please refer to Figure 6 , Figure 6 This is a schematic diagram of a device for analyzing the behavior of a convolutional neural network in regression analysis.

[0114] This embodiment provides a regression analysis convolutional neural network behavior parsing device, including:

[0115] The first processing module 601 is used to acquire the input image and the regression analysis convolutional neural network to be parsed, and to input the input image into the regression analysis convolutional neural network for processing to obtain the desired output value.

[0116] The generation module 602 is used to create a population matrix in a random initialization manner, where each element in the population matrix is ​​a binary filter for the behavior analysis of convolutional neural networks in regression analysis.

[0117] The iteration module 603 is used to update the population matrix based on a preset number of iterations and determine the optimal element individual from the updated population matrix according to preset screening conditions.

[0118] The second processing module 604 is used to perform a convolution operation between the binary filter corresponding to the optimal element individual and the input image to obtain the target image, and then input the target image into the regression analysis convolutional neural network for processing to obtain the target output value.

[0119] The parsing module 605 is used to determine the relationship between the input and output values ​​of the regression analysis convolutional neural network based on the first difference and the second difference, so as to analyze the prediction behavior of the regression analysis convolutional neural network. The first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

[0120] It will be understood by those skilled in the art that all or some of the steps and apparatuses in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. As is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0121] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0122] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement any of the above regression analysis convolutional neural network behavior parsing methods.

[0123] refer to Figure 7 , Figure 7 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0124] The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0125] The memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store operating devices and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701 to execute the regression analysis convolutional neural network behavior parsing method of the embodiments of this application.

[0126] The input / output interface 703 is used to implement information input and output;

[0127] The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0128] Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704);

[0129] The processor 701, memory 702, input / output interface 703, and communication interface 704 are connected to each other within the device via bus 705.

[0130] It is understood that the content of the above method embodiments is applicable to the embodiments of this electronic device. The specific functions implemented by the embodiments of this electronic device are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0131] This application also provides a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement the regression analysis convolutional neural network behavior parsing method as described in any of the above specific embodiments.

[0132] This application also discloses a computer program product, including a computer program or computer instructions, which are stored in a computer-readable storage medium. The processor of a computer device reads the computer program or computer instructions from the computer-readable storage medium and executes the computer program or computer instructions, causing the computer device to perform the regression analysis convolutional neural network behavior parsing method as described in any of the preceding embodiments.

[0133] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0134] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0135] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or units, and may be electrical, mechanical, or other forms.

[0136] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0137] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0138] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] Although the description of this application has been quite detailed and particularly focused on several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment. Rather, it should be considered as effectively covering the intended scope of this application by referring to the appended claims and taking into account the broad possible interpretations of these claims provided by the prior art. Furthermore, the foregoing description of this application with respect to embodiments foreseeable by the inventors is intended to provide a useful description, and non-substantial modifications to this application that have not yet been foreseen may still represent equivalent modifications.

Claims

1. A method for analyzing the behavior of convolutional neural networks in regression analysis, characterized in that, The method includes: The input image and the regression analysis convolutional neural network to be parsed are obtained, and the input image is input into the regression analysis convolutional neural network for processing to obtain the desired output value; A population matrix is ​​created using random initialization, where each element in the population matrix is ​​a binary filter for behavioral analysis of the regression analysis convolutional neural network. The population matrix is ​​updated based on a preset number of iterations, and the optimal element individual is determined from the updated population matrix according to preset screening conditions. The binary filter corresponding to the optimal element is convolved with the input image to obtain the target image, and the target image is then input into the regression analysis convolutional neural network for processing to obtain the target output value. The relationship between the input and output values ​​of the regression analysis convolutional neural network is determined based on the first difference and the second difference to analyze the prediction behavior of the regression analysis convolutional neural network. The first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

2. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 1, characterized in that, The step of updating the population matrix based on a preset number of iterations and determining the optimal individual from the updated population matrix according to preset screening conditions includes: Update the number of individuals and the characteristics of individuals in the current population matrix; Determine the fitness of each individual element in the updated current population matrix; Based on the fitness of each individual element, multiple selected individual elements are chosen from the current population matrix; A new population matrix is ​​generated based on the multiple selected element individuals; The new population matrix is ​​used as the current population matrix, and the process returns to the step of updating the number of individual elements and the characteristics of individual elements in the current population matrix. When the number of times the new population matrix is ​​generated is equal to the preset number of iterations, the new population matrix is ​​used as the updated population matrix, and the element with the best fitness in the updated population matrix is ​​used as the optimal element.

3. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 2, characterized in that, Determining the fitness of each individual element in the updated current population matrix includes: The input image is convolved with each binary filter in the population matrix to obtain images after multiple convolution operations; The image after each convolution operation is input into the regression analysis convolutional neural network for processing to obtain the actual output value corresponding to each image after the convolution operation; The expected output value is compared with each of the actual output values ​​to determine the absolute error value corresponding to each of the binary filters, wherein the absolute error value is used as the fitness of the individual element.

4. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 2, characterized in that, The step of selecting multiple selected element individuals from the current population matrix based on the fitness of each element individual includes: Obtain the first candidate element individual and the second candidate element individual from the current population matrix; Determine whether the fitness of the first candidate element is greater than that of the second candidate element. If yes, then the second candidate element is selected as the filtered element. If no, then the first candidate element is selected as the filtered element. Return to the step of obtaining the first candidate element individual and the second candidate element individual from the current population matrix, until the number of the filtered element individuals is equal to the preset number threshold.

5. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 2, characterized in that, The update of the number of individuals and the characteristics of individuals in the current population matrix includes: The current population matrix is ​​divided into multiple element groups, and each element group includes a first element individual and a second element individual; Based on a first preset probability value, a target element group is determined from the plurality of element groups, and a plurality of third element individuals are generated according to the characteristics of the first element individual and the characteristics of the second element individual in each target element group. Based on the second preset probability value, a mutation operation is performed on each of the third element individuals to obtain a new element individual; The updated current population matrix is ​​generated based on the new element individuals.

6. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 5, characterized in that, The step of determining a target element group from the plurality of element groups based on a first preset probability value, and generating a plurality of third element individuals based on the characteristics of the first element individuals and the characteristics of the second element individuals within each target element group, includes: Based on a first preset probability value, a target element group is determined from the plurality of element groups; The first and second individual elements in the target element group are respectively converted into a first vector and a second vector, and both the first vector and the second vector include multiple parameters. At least one parameter in the first vector is swapped with the target parameter in the second vector, and the first and second vectors after the parameter swap are respectively converted into the third element individual.

7. The regression analysis method for analyzing the behavior of convolutional neural networks according to claim 5, characterized in that, The step of performing a mutation operation on each of the third element individuals based on a second preset probability value to obtain a new element individual includes: The third element individual is converted into a third vector, which includes multiple parameters; Based on the second preset probability value, the value of the target parameter among the multiple parameters is adjusted to obtain the third vector after parameter adjustment and convert it into the new element individual.

8. A device for analyzing the behavior of a convolutional neural network in regression analysis, characterized in that, The device includes: The first processing module is used to acquire the input image and the regression analysis convolutional neural network to be parsed, and to input the input image into the regression analysis convolutional neural network for processing to obtain the desired output value; A generation module is used to create a population matrix in a random initialization manner, wherein each element in the population matrix is ​​a binary filter for the behavior analysis of the regression analysis convolutional neural network. An iteration module is used to update the population matrix based on a preset number of iterations and determine the optimal element individual from the updated population matrix according to preset screening conditions. The second processing module is used to perform a convolution operation between the binary filter corresponding to the optimal element individual and the input image to obtain the target image, and input the target image into the regression analysis convolutional neural network for processing to obtain the target output value; The parsing module is used to determine the relationship between the input and output values ​​of the regression analysis convolutional neural network based on a first difference and a second difference, so as to analyze the prediction behavior of the regression analysis convolutional neural network, wherein the first difference is the difference between the target image and the input image, and the second difference is the difference between the target output value and the expected output value.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the regression analysis convolutional neural network behavior analysis method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the regression analysis convolutional neural network behavior parsing method as described in any one of claims 1 to 7.

Images