Image recognition method based on deep reinforcement learning and evolutionary algorithm

Abstract

The application discloses an image recognition method based on deep reinforcement learning and an evolutionary algorithm, and comprises the following steps: S1, collecting image data and performing standard processing on the image; S2, defining an optimization target and an optimization target function; S3, solving a multi-target optimization function F based on deep learning and the evolutionary algorithm; S4, performing optimization training on the model based on the deep learning and the evolutionary algorithm; and S5, classifying image data by using the optimized image recognition model. The application can be applied to image recognition, and the model structure is optimized, so that the image recognition precision is effectively improved, and the image classification error rate is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image recognition, in particular to an image recognition method based on deep reinforcement learning and evolutionary algorithm, and the image recognition is an application of deep learning algorithm in practice, and has strong application value in real life. BACKGROUND

[0002] With the advent of the information age and the rapid development of machine learning, deep learning is gradually widely used in various fields in real life, such as computer vision, data mining, natural language processing and multimedia learning, and has achieved many proud achievements, and image recognition is also an important application of deep learning.

[0003] Image recognition refers to processing, analyzing and understanding images by computer to identify various different patterns of targets and objects, and is a practical application of deep learning algorithm. The accuracy of various image recognition methods is highly dependent on the performance of the neural network used, and the structure of the neural network is an important factor affecting the complexity and classification error rate of the neural network, thereby affecting the performance of the neural network. Therefore, optimizing the structure of the neural network is an important means to improve the image recognition method, and the existing image recognition method does not pay enough attention to the structure optimization of the neural network in the recognition model. SUMMARY

[0004] The present application is to solve the above-mentioned problems of the prior art, and proposes an image recognition method based on deep reinforcement learning and evolutionary algorithm, so as to optimize the neural network structure used in the image recognition model to improve the accuracy of image recognition, thereby reducing the error rate of image recognition classification.

[0005] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows:

[0006] The image recognition method based on deep reinforcement learning and evolutionary algorithm of the present application is characterized by the following steps:

[0007] Step 1, collect image data and perform standardization processing;

[0008] Obtain T image samples and their class labels, and according to the class label of each image sample, extract the attribute features corresponding to each image sample, thereby obtaining the attribute feature set of the image sample Wherein, x t represents the attribute feature of the tth image sample, y t represents the true class label of the tth image sample, t = 1, 2, …, T; T represents the number of image samples;

[0009] Step 2, define the optimization target and the optimization target function:

[0010] Step 2.1, constructing an image recognition network Model, including: an input layer, a hidden layer and an output layer; and using formula (1) to construct a decision vector C of the image recognition network Model:

[0011] C = (p1, p2,..., p i ,...,p k ,a1,a2,...,a i ,...,a k ,r,z) (1)

[0012] In formula (1), p i is the number of neurons in the i-th hidden layer of the image recognition network Model, a i is the activation function selected for the i-th hidden layer of the image recognition network Model, z is the optimizer of the image recognition network Model, and r is the learning rate of the image recognition network Model; 1≤i≤k; k represents the total number of hidden layers in the image recognition network Model;

[0013] Step 2.2, taking the minimum recognition error rate of the image recognition network Model for T image samples as a first optimization objective function f1; taking the minimum number of neurons in the hidden layer of the image recognition network as a second optimization objective function f2; thereby establishing a multi-objective optimization function

[0014] Step three, solving the multi-objective optimization function F based on deep learning and evolutionary algorithm;

[0015] Step 3.1, defining the upper bound of the decision vector of the optimization objective as Upper; defining the lower bound of the decision vector of the optimization objective as Lower; defining the population as P, defining the size of the population P as N; defining the uniformly distributed weight vector as w; defining the neighbor index of each weight vector as B; defining the deep neural network as Q, defining the experience replay pool as U, defining the candidate operator set as OP; defining the current generation as gen, and initializing gen = 1; defining the total number of generations as G;

[0016] Step 3.2, initializing the gen-th generation population P wherein, represents the n-th individual in the gen-th generation population P gen ;

[0017] Defining each individual in the gen-th generation population P gen is composed of the decision vector C and the multi-objective optimization function F;

[0018] Step 3.3: Initialize n = 1; randomly initialize the nth individual in the population P of the gen-th generation gen in the decision vector

[0019] Step 3.4: Use the decision vector of the nth individual in the population P of the gen-th generation gen to construct the tth image recognition network in the population P of the gen-th generation and input the attribute features of T image samples into for processing and obtain the predicted class label set ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ combination is state state input into the deep neural network Q, and the probability value of each operator in the mth candidate operator set of the gen generation is obtained , and an operator op is selected from the probability value through roulette;

[0025] Step 3.9, cross mutation is performed on the parent population P and , thereby generating offspring P

[0026] Step 3.10, an adjacent neighbor y is selected from the neighbor index , the aggregate function values of the offspring P and the adjacent neighbor y are calculated respectively, if the aggregate function value of the offspring P is less than that of the adjacent neighbor y, the offspring P replaces the adjacent neighbor y, otherwise, no replacement is made; thereby all the adjacent neighbors in the neighbor index are traversed, and the population P gen+1 of the (gen+1)th generation is obtained;

[0027] Step 3.11, the promotion amplitude of the aggregate function value of the offspring P relative to the aggregate function value of the parent P is defined as the reward , and the decision vector and the weight vector of the offspring P are combined into a state state2, and , op, state2 are taken as a set of training data, and are put into the experience replay pool U;

[0028] Step 3.12, a plurality of sets of training data are randomly selected from the experience replay pool T and input into the deep neural network Q for training, a trained deep neural network is obtained, and is assigned to Q;

[0029] Step 3.13, gen+1 is assigned to gen, and whether gen G is established is judged, if yes, the step 3.3 is returned for sequential execution, otherwise, the population P G of the Gth generation is output;

[0030] Step four, the image samples are classified;

[0031] Let P represent the jth individual of the final population P G , and j=1 is initialized;

[0032] Step 4.1, the model parameters of the image recognition method model Model are replaced by the decision vector of the jth individual P ; Thus, the image recognition method model is obtained. j ;

[0033] Step 4.2: Transfer attribute features Input the j-th image recognition model. j The images are processed to obtain T image recognition results. Where y” t The image recognition model is represented by Model. j The predicted class label on the sample of the t-th image, based on the recognition result. and the real category label set Construct the j-th image recognition model j Loss function j ;

[0034] Step 4.3: Use gradient descent to model the j-th image recognition model. j Perform training and calculate the loss function. j To update model parameters until Loss j The process continues until convergence is achieved, thus obtaining the j-th trained image recognition model. j ;

[0035] Step 4.4: Transfer attribute features Input the j-th image recognition model after training. j The image is then processed to obtain the updated image recognition result. in, This represents the j-th image recognition model after training. j The predicted category label for the t-th image sample is based on the updated image recognition result. With the real label information Loss function value between j Calculate the population P of generation G. G The j-th individual The two optimization objective function values;

[0036] Step 4.5: Assign j+1 to j. If j>N, it means that the population P of generation G is... G Once the two objective function values ​​for all individuals have been optimized, proceed to step 4.6; otherwise, return to step 4.1 and execute sequentially.

[0037] Step 4.6: In a two-dimensional coordinate system, plot the population P of generation G. GThe two optimized objective function values of each individual are fitted to obtain a fitting curve, the decision vector of the individual corresponding to the inflection point of the fitting curve is selected, and the decision vector is substituted into the image recognition network Model, so that the optimal image classification model Model is obtained final For identifying and classifying images.

[0038] The electronic device comprises a memory and a processor, wherein the memory is configured to store a program supporting the processor to execute the image recognition method, and the processor is configured to execute the program stored in the memory.

[0039] The computer readable storage medium stores a computer program, and the computer program is configured to execute the steps of the image recognition method when executed by a processor.

[0040] Compared with the prior art, the present application has the following beneficial effects:

[0041] 1. The present application proposes a multi-objective operator selection method based on reinforcement learning, which uses a deep neural network to learn a strategy by taking neural network structure parameters as states, candidate operators as actions, fitness improvement as rewards, and population evolution as the environment, and estimates the Q value of each action under a given state. The Q value represents the cumulative fitness improvement brought by an operator in the future rather than in the past, so it is expected to produce better offspring solutions in future generations. Therefore, the image recognition method model structure is optimized, and the accuracy of image classification and recognition is improved.

[0042] 2. The proposed operator selection method is embedded into a decomposition-based multi-objective evolutionary algorithm with dynamic resource allocation and applied to the image recognition method to improve the accuracy of the image recognition method. In the proposed method, the agent iteratively updates the deep neural network to guide the selection of operators, and by using four different types of mutation operators as candidate operators, the proposed method has high universality and achieves good classification and recognition performance in experiments. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a complete flowchart of the inventive method.

[0044] Figure 2 It is a flowchart of the multi-objective optimization algorithm.

[0045] Figure 3 It is a process diagram of reinforcement learning. DETAILED DESCRIPTION

[0046] In this embodiment, as Figure 1As shown in the image recognition method based on deep reinforcement learning and evolutionary algorithm is performed according to the following steps:

[0047] Step 1, collect image data and perform standardization processing;

[0048] Obtain T image samples and their class labels, and according to the class label of each image sample, extract the attribute features corresponding to each image sample, thereby obtaining the attribute feature set of the image sample Wherein, x t represents the attribute feature of the tth image sample, y t represents the true class label of the tth image sample, t=1,2,...,T; T represents the number of image samples;

[0049] Step 2, define the optimization target and the optimization target function:

[0050] Step 2.1, construct the image recognition network Model, including: input layer, hidden layer and output layer; and use formula (2) to construct the decision vector C of the image recognition network Model:

[0051] C=(p1,p2,...,p i ,...,p k ,a1,a2,...,a i ,...,a k ,r,z) (2)

[0052] In formula (1), p i is the number of neurons in the i-th hidden layer of the image recognition network Model, a i is the activation function selected in the i-th hidden layer of the image recognition network Model, z is the optimizer of the image recognition network Model, r is the learning rate of the image recognition network Model; 1≤i≤k; k represents the total number of hidden layers in the image recognition network Model

[0053] Step 2.2, the first optimization target function f1 is that the recognition error rate of the image recognition network Model for T image samples is minimum; the second optimization target function f2 is that the number of neurons in the hidden layer of the image recognition network is minimum; thereby establishing a multi-objective optimization function

[0054] Step 3, as shown in the image recognition method based on deep reinforcement learning and evolutionary algorithm is solved by using the multi-objective optimization function F based on deep learning and evolutionary algorithm; Figure 2

[0055] ​Step 3.1, define the upper bound of the decision vector of the optimization objective as Upper; define the lower bound of the decision vector of the optimization objective as Lower; define the population as P, define the size of the population P as N; define the uniformly distributed weight vector as w; define the neighbor index of each weight vector as B; define the deep neural network as Q, define the experience replay pool as U, define the candidate operator set as OP; define the current generation as gen, and initialize gen = 1; define the total number of generations as G;

[0056] The following are four candidate operators in the candidate operator set OP:

[0057] The simulated binary crossover operator is shown in equation (3):

[0058]

[0059]

[0060] In equation (3), x = (x1,...,x d ,...,x D ) is the offspring solution, D is the dimension of the decision variable, y1, y2 are two parents, μ is a random value uniformly sampled in the range [0, 1], η is a predefined parameter, usually taking 1.

[0061] The crossover operator in MOEA / D-M2M is shown in equation (4):

[0062]

[0063] In equation (4), r1, r2 are random values uniformly sampled in the range [0, 1], gen is the current evolution generation, and maxgen is the maximum evolution generation.

[0064] DE / rand / 1 and DE / rand / 2 two difference operators:

[0065] The DE / rand / 1 operator is shown in equation (5):

[0066]

[0067] The DE / rand / 2 operator is shown in equation (6):

[0068]

[0069] where F and CR are predefined parameters, and r is a random variable sampled in the range [0, 1].

[0070] Step 3.2, initialize the gen generation population according to the upper bound Upper and the lower bound Lower where, denote the \(gen\)-th generation population \(P\) gen the \(n\)-th individual in it;

[0071] Define each individual in the \(gen\)-th generation population \(P\) gen is composed of a decision vector \(C\) and a multi-objective optimization function \(F\);

[0072] Step 3.3, initialize \(n = 1\); randomly initialize the \(n\)-th individual gen in the \(gen\)-th generation population \(P\) decision vector of

[0073] Step 3.4, use the decision vector of the \(n\)-th individual gen in the \(gen\)-th generation population \(P\) to construct the \(t\)-th image recognition network in the \(gen\)-th generation population \(P\) gen and input the attribute features of \(T\) image samples into for processing and obtain the predicted class label set where, denote the predicted class label on the attribute feature \(x\) of the \(t\)-th image sample ; t

[0074] Step 3.5, based on and the true class label set construct the loss function value of and calculate the multi-objective optimization function of

[0075] Step 3.6, assign \(t + 1\) to \(t\), judge whether \(t < N\) holds, if it holds, return to Step 3.4 to execute sequentially, otherwise, it means that each individual in the \(gen\)-th generation population \(P\) gen has completed the learning of the attribute features and execute Step 3.7;

[0076] Step 3.7, mating pool selection:

[0077] Select \(M\) individuals from the \(gen\)-th generation population \(P\) gen through the tournament algorithm to form the \(gen\)-th generation mating pool where, denote the \(m\)-th parent in the \(gen\)-th generation mating pool Parent; \(M\) is the number of parents; gen

[0078] Step 3.8: Calculate the m-th parent generation. weight vector and from Neighbor index Randomly select a neighbor as the second parent. Will Decision vector and weight vector Combining into states State The input is fed into the deep neural network Q to obtain the set of candidate operators for the gen-th generation m. The probability value of each operator is determined, and an operator op is selected from the probability values ​​by roulette wheel selection;

[0079] Step 3.9: Use operator op to pair and Crossover mutation is performed to produce offspring.

[0080] Step 3.10, Index from Neighbors Select a neighbor y and calculate the offspring of each. The aggregation function value of the neighbor y, if the child If the aggregation function value of y is less than the aggregation function value of its neighbor y, then the child y is used. Replace neighbor y; otherwise, do not replace it; thus, the neighbor index is traversed. All neighbors are identified, and the population P of generation gen+1 is obtained. gen+1 The definition of the aggregation function is shown in equation (7):

[0081]

[0082] In equation (7), w = (w1, ..., w M ) is the weight vector corresponding to the solution x. The ideal point is composed of the minimum target value in the population. The generated offspring solution x is compared with all solutions in its parent's neighborhood, and its fitness improvement for the neighborhood individual y is calculated. The calculation formula is shown in Equation (8):

[0083]

[0084] In equation (8), w is the weight vector of y. To avoid the dominant solution reducing the diversity of the subpopulation within the neighborhood, a sub-solution can replace at most n subpopulations within the neighborhood. r There are one solution, therefore the relative fitness improvement rate (NFI) within the neighborhood. x Further calculations are performed to replace n with the child solution x. r The sum of the fitness improvements of each neighborhood solution.

[0085] Step 3.11, define offspring the aggregate function value of the parent the promotion amplitude of the aggregate function value of the offspring as a reward At the same time, the decision vector and the weight vector of the offspring are combined into a state state2, and op, state2is a set of training data, and is put into the experience replay pool U. Considering the randomness of the evolutionary algorithm, the method of the application applies the operator op and the neighborhood fitness improvement rate NFI obtained in the historical period x is constructed into a tuple (op, NFI x ) and saved to a first-in-first-out queue R. The calculation of the operator reward is shown in formula (9):

[0086]

[0087] Step 3.12, randomly select several sets of training data from the experience replay pool T and input them into the deep neural network Q for training, obtain the trained deep neural network, and assign it to Q. As shown in Figure 3 , the value-based method processes the discrete action space by fitting an action-value function, which calculates the expected cumulative reward (Q value) when taking action a t in state s t . Specifically, the action-value function is defined as follows:

[0088]

[0089] The method of DeepQ-Network (DQN) uses a deep neural network to approximate the state-value function with continuous action space, and the input and output of the neural network are the current state and the Q value that can be obtained by executing all actions in the state, respectively. DQN stores the obtained tuples in the experience replay pool as a training set, and crosses the action with the training of the neural network. The loss function of the neural network uses mean square error to represent, and uses the gradient descent method to reduce the loss.

[0090]

[0091] In formula (11), T is the training set, Q(s t ,a t ) represents the output of the a t th neuron in the neural network when the input vector s t , and q t is the Q value of executing action a t in state s t .

[0092]

[0093] Expected output q t not only includes the current reward r t , but also includes the maximum reward for performing the next operation Due to the recursive nature of the formula, q t can represent the maximum cumulative future reward.

[0094] Step 3.13: Assign gen + 1 to gen, and determine whether gen < G holds. If it holds, return to Step 3.3 and execute sequentially. Otherwise, output the Gth generation population P G ;

[0095] Step Four: Classify the image samples. Let represent the jth individual of the final population P G , and initialize j = 1;

[0096] Step 4.1: Replace the model parameters of the image recognition method model Model with the decision vector of the jth individual to obtain the image recognition method model Model ; j ;

[0097] Step 4.2: Input the attribute features into the jth image recognition model Model j for processing to obtain T image recognition results where y”<00,00080>represents the predicted class label of the jth image recognition model Model j on the sample of the tth image. According to the recognition results and the true class label set construct the loss function Loss of the jth image recognition model Model j ; j ;

[0098] Step 4.3: Use the gradient descent method to train the jth image recognition model Model j and calculate the loss function Loss j to update the model parameters until Loss j converges, thereby obtaining the trained jth image recognition model Model j ;

[0099] Step 4.4: Input the attribute features into the trained jth image recognition model Model j for processing to obtain the updated image recognition results wherein, represents the jth image recognition model Model j the predicted class label of the tth image sample according to the updated image recognition result and the real label information loss function value Loss' j , the two optimization objective function values of the jth individual in the Gth generation population P G

[0100] Step 4.5, j+1 is assigned to j, if j>N, it indicates that the two optimization objective function values of all individuals in the Gth generation population P G are optimized, and step 4.6 is executed, otherwise, step 4.1 is returned to be sequentially executed;

[0101] Step 4.6, the two optimization objective function values of each individual in the Gth generation population P G are fitted in a two-dimensional coordinate system to obtain a fitting curve, the decision vector of the individual corresponding to the inflection point of the fitting curve is selected, and the image recognition network Model is substituted into the decision vector, thereby obtaining the optimal image classification model Model final for identifying and classifying images.

[0102] In this embodiment, an electronic device includes a memory for storing a program supporting a processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0103] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program is run by a processor to execute the steps of the above method.​​

Claims

1. An image recognition method based on deep reinforcement learning and evolutionary algorithm, characterized in that The steps are as follows: Step one, collect image data and perform standardization processing; T image samples and their class labels are acquired, and attribute features corresponding to each image sample are extracted according to the class label of each image sample, so as to obtain an attribute feature set of the image sample wherein x t represents the attribute feature of the tth image sample, y t represents the true class label of the tth image sample, t = 1, 2, …, T; T represents the number of image samples; Step two, define optimization target and optimization objective function: Step 2.1, construct image recognition network Model, including: input layer, hidden layer and output layer; and use formula (1) to construct the decision vector C of the image recognition network Model: C=(p1,p2,...,p i ,...,p k ,a1,a2,...,a i ,...,a k ,r,z) (1) In formula (1), p i is the number of neurons of the i-th hidden layer in the image recognition network Model, a i is the activation function selected for the i-th hidden layer in the image recognition network Model, z is the optimizer of the image recognition network Model, r is the learning rate of the image recognition network Model; 1≤i≤k; k represents the total number of hidden layers in the image recognition network Model. Step 2.2, taking the minimum recognition error rate of the image recognition network Model on T image samples as a first optimization objective function f1, and taking the minimum number of neurons in the hidden layer of the image recognition network as a second optimization objective function f2, thereby establishing a multi-objective optimization function Step three, solve the multi-objective optimization function F based on deep learning and evolutionary algorithm; Step 3.1, define the upper bound of the decision vector of the optimization target as Upper; define the lower bound of the decision vector of the optimization target as Lower; define the population as P, define the size of the population P as N; define the uniformly distributed weight vector as w; define the neighbor index of each weight vector as B; define the deep neural network as Q, define the experience replay pool as U, define the candidate operator set as OP; define the current generation as gen, and initialize gen=1; define the total number of generations as G; Step 3.

2. Initialize the population of the gen-th generation according to the upper bound Upper and the lower bound Lower wherein, denotes the population P of the gen-th generation gen the n-th individual in the population. Definition of the population P of the gen-th generation gen Each individual in P is composed of a decision vector C and a multi-objective optimization function F; Step 3.3, initialize n = 1; randomly initialize the first generation population P gen of the nth individual in P the decision vector of the nth individual in P Step 3.4: Utilize the population P of generation gen gen The nth individual Decision vector Construct the gen-th generation population P gen The t-th image recognition network And the attribute features of T image samples enter Process it and obtain Predicted category label set in, express The attribute features x of the t-th image sample t The predicted category label; Step 3.5, based on with the real set of class labels construct the loss function value and compute the multi-objective optimization function Step 3.6, assign t+1 to t, judge whether t<N is true, if true, return to step 3.4 for sequential execution, otherwise, it means that the population P in the gen generation is completed gen Each individual in the population P in the gen generation completes the learning of the attribute characteristics of the population P in the gen generation, and executes step 3.7; Step 3.7, mating pool selection: M individuals are selected from the population P of the gen-th generation by a tournament algorithm to form a mating pool of the gen-th generation gen M individuals are selected from the population P of the gen-th generation by a tournament algorithm to form a mating pool of the gen-th generation wherein, the m-th parent in the mating pool Parent of the gen-th generation gen M is the number of parents; Step 3.8: Calculate the m-th parent generation. weight vector and from Neighbor Index Randomly select a neighbor as the second parent. Will Decision vector and weight vector Combining into states State The input is fed into the deep neural network Q to obtain the set of candidate operators for the gen-th generation m. The probability value of each operator is determined, and an operator op is selected from the probability values ​​by roulette wheel selection; Step 3.9, cross variation with the operator op on and to produce offspring Step 3.10, select a neighbor y from the neighbor index Compute the offspring and the neighbor y's aggregate function value, if the offspring 's aggregate function value is less than the neighbor y's aggregate function value, replace the neighbor y with the offspring , otherwise, do not replace; thus, traverse all the neighbors in the neighbor index and get the gen+1 generation population P gen+1 ; Step 3.11, Define offspring the aggregate function value of the parent the magnitude of the boost in the aggregate function value of the offspring as a reward simultaneously combine the decision vector and the weight vector of the offspring into a state state2, and op, state2 as a set of training data, and put it into the experience replay pool U; Step 3.12, randomly select several groups of training data from the experience replay pool T and input them into the deep neural network Q for training, obtain the trained deep neural network, and assign it to Q; Step 3.13, assign gen+1 to gen, judge whether gen G ; Step four, classify the image samples; Let denote the jth individual of the final population P G , initialize j = 1; Step 4.1, replace the model parameters of the image recognition method model Model with the jth individual 's decision vector Thus obtaining the image recognition method model Model j ; Step 4.2, attribute features Input the jth image recognition model Model j , and obtain T image recognition results Where y'' is t the predicted class label of the jth image recognition model Model j on the sample of the tth image, according to the recognition result and the real class label set , the loss function Loss of the jth image recognition model Model j is constructed j ; Step 4.3, identifying the jth image recognition model Model j using gradient descent method, training and calculating loss function Loss j to update model parameters until Loss j converges, thereby obtaining the trained jth image recognition model Model j ; Step 4.4: Transfer attribute features Input the j-th image recognition model after training. j The image is then processed to obtain the updated image recognition result. in, This represents the j-th image recognition model after training. j The predicted category label for the t-th image sample is based on the updated image recognition result. With real label information Loss function value between j Calculate the population P of generation G. G The j-th individual The two optimization objective function values; Step 4.5, j+1 is assigned to j, if j>N, it means that the optimization of all individual's two optimization objective function values in the Gth generation population P is completed, and step 4.6 is executed, otherwise, return to step 4.1 for sequential execution; G Step 4.6, the two optimization objective function values of all individuals in the Gth generation population P are optimized, and step 4.7 is executed, otherwise, return to step 4.1 for sequential execution; Step 4.6, fitting two optimization objective function values of each individual in the Gth generation population P in a two-dimensional coordinate system to obtain a fitting curve, selecting a decision vector of an individual corresponding to an inflection point of the fitting curve, and substituting the decision vector into the image recognition network Model to obtain an optimal image classification model Model G Step 4.6, fitting two optimization objective function values of each individual in the Gth generation population P in a two-dimensional coordinate system to obtain a fitting curve, selecting a decision vector of an individual corresponding to an inflection point of the fitting curve, and substituting the decision vector into the image recognition network Model to obtain an optimal image classification model Model final for identifying and classifying images.

2. An electronic device comprising a memory and a processor, characterized in that The memory is used to store the program supporting the processor to execute the image recognition method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to perform the steps of the image recognition method of claim 1.

Images