Multi-scale skin anomaly classification method based on flattened genetic programming individual structure

Abstract

The invention discloses a multi-scale skin anomaly classification method based on a flattened genetic programming individual structure. The method comprises the steps that 1, a skin sample image data set is collected and preprocessed; 2, according to the defined function set, the terminal set and the designed GP individual tree structure, using an evolutionary algorithm framework DEAP to generate GP individuals and constructing an initial population; 3, evolving the GP individuals in the population by adopting an evolutionary learning method, and searching an optimal GP individual after iterative evolution; and 4, using the optimal GP individual as a feature extraction classification model, inputting a skin sample image data set to be classified into the feature extraction classification model to extract feature vectors, and finishing final classification of skin sample images by using a linear SVM classifier. According to the method, the optimal feature combination scheme suitable for skin image classification can be evolved, and the skin image features under different scales can be extracted and combined, so that the anomaly classification of the skin image can be realized under the condition that expert domain knowledge is not needed.

Description

Technical Field

[0001] The present invention belongs to the field of image processing in evolutionary algorithms / computer vision, and specifically is a multi-scale skin abnormality classification method based on flattened genetic programming individual structure. Background Art

[0002] To examine skin abnormalities, dermatologists usually follow the ABCD (asymmetry, border irregularity, color change, and dermoscopic structure) rule of dermoscopy, which calculates scores by measuring these four abnormal characteristics to effectively classify different types of skin abnormality images.

[0003] These important visual properties and the accessibility of a large number of skin abnormality images have made many researchers interested in proposing CAD frameworks to help dermatologists in early identification. Along these lines, it is necessary to define strategies that can capture informative features such that the captured features are suitable in some way to simulate these clinical properties and can utilize different texture, color, local and global features. Local features are extracted from sub-images, while global features are extracted from the entire image. This classification method is potentially useful not only to quickly and accurately classify a type of skin abnormality in practical situations, but also to effectively identify important features to help dermatologists understand the key abnormality characterization from skin abnormalities.

[0004] Genetic programming (GP), as a computational method that simulates the biological evolution process, has been widely used in feature learning in image classification tasks. GP is a population-based algorithm in which each individual in the population (also called a program or tree) represents a candidate solution to the problem. GP is based on a tree-like program structure that can integrate multiple tasks, perform feature extraction and construction at the same time, and continuously iterate and optimize during the evolution process, thereby effectively solving image classification problems. The GP method usually uses a tree structure to dynamically construct and optimize individuals to extract and combine key features of the image. It has good interpretability and has unique advantages in processing small sample data and tasks that require high model interpretability.

[0005] Therefore, GP is very suitable for the identification and classification of skin abnormality images. However, in most GP-based image classification methods, the deep and wide individual tree structure often makes the individuals too complex. Although this complex structure can improve the classification performance, it may also introduce redundant features, thereby affecting the effectiveness of feature learning and reducing the interpretability of the GP individual model. In addition, existing GP methods usually only extract features from the original image or local area, and fail to fully utilize the detailed information of the image at different scales, which further limits the improvement of the individual image classification performance of the GP program structure. Summary of the invention

[0006] The present invention aims to address the deficiencies of the above-mentioned prior art and proposes a multi-scale skin abnormality classification method based on a flattened genetic programming individual structure, in order to identify and extract skin image features at different scales, thereby achieving accurate abnormality classification of skin images without the need for expert domain knowledge.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0008] The multi-scale skin abnormality classification method based on flattened genetic programming individual structure of the present invention is characterized in that it comprises the following steps:

[0009] Step 1: Obtain skin image samples with real labels and convert them into grayscale images to construct a skin grayscale image training set ,in, represents the i-th labeled skin grayscale image, express The true category label, express The total number of medium skin grayscale images;

[0010] Obtain skin image samples to be classified and convert them into grayscale images to construct a skin grayscale image test set , represents the jth unlabeled skin grayscale image, represents the total number of skin grayscale images in the test set;

[0011] Step 2: Based on the evolutionary algorithm framework Generate GP individuals and construct the initial population;

[0012] Step 3: Evolve the g-th generation population using evolutionary learning methods GP individuals in, find the global best GP individual;

[0013] Step 4: The individual tree structure corresponding to the global best GP individual is used as the optimal skin image classification model, and is used to process the test set T to obtain the feature vector of T After that, the linear SVM classifier is input for classification, and we get The predicted classification label of Indicates T The feature vector of .

[0014] The multi-scale skin abnormality classification method based on flattened genetic programming individual structure described in the present invention is also characterized in that step 2 comprises the following steps:

[0015] Step 2.1: Build a function set ;

[0016] The function set Contains: Area resampling function , feature extraction function And feature combination function ;

[0017] The area resampling function Includes: Global resampling functions and the local resampling function ;

[0018] The global resampling function Includes: Downsampling function , upsampling function and the clipping function ;in, Used to perform one or two upsampling operations on the skin grayscale image. Used to perform one or two downsampling operations on the skin grayscale image, cropping function Used to adjust the input size of skin grayscale images;

[0019] The local resampling function Includes: Square area resampling function and rectangular area resampling function ;in, It is used to select a square area from the skin grayscale image and perform an upsampling operation. Used to select a rectangular area from the skin grayscale image and perform an upsampling operation;

[0020] The feature extraction function It is composed of several image feature extraction functions;

[0021] The feature combination function FC is composed of several image feature combination functions, which is used to combine the feature extraction function The extracted features are combined into a feature vector;

[0022] Step 2.2: Construct a terminal set Terminals to define the parameters of related functions in the function set; the terminal set Includes: skin grayscale image, horizontal and vertical coordinate parameters of the upper left corner rectangular area of ​​R_Up_Sample in LRS, horizontal and vertical coordinate parameters of the upper left corner of the square area in S_Up_Sample, and the size of the rectangular area , Long parameters for adjusting skin grayscale images in and width parameters ;

[0023] Step 2.2: Set the maximum number of iterations , the current number of iterations is ,initialization ;

[0024] Step 2.3: Based on the function set Funcs and the terminal set Terminals, use Algorithm generation Individual trees as GP individuals, thus forming the g-th generation population ;

[0025] The structure of constructing any individual tree includes: input layer, region resampling layer, feature extraction layer, feature combination layer and output layer;

[0026] The generation method of any individual tree is as follows: select several functions from the regional resampling function RS as nodes of the regional resampling layer in the individual tree; select several functions from the feature extraction function FE as nodes of the feature extraction layer in the individual tree, and connect them with the nodes of the regional resampling layer; select a function from the feature combination function FC as a node of the feature combination layer in the individual tree, and connect it with the nodes of the feature extraction layer; according to the terminal set Define the parameters of the input layer and regional resampling layer in the individual tree, thereby obtaining an individual tree; wherein:

[0027] The input layer is used to input a skin grayscale image;

[0028] The regional resampling layer uses the selected regional resampling function The function in resamples the skin grayscale image or the regional image in the skin grayscale image input by the input layer;

[0029] The feature extraction layer uses the selected feature extraction function The function in extracts features from the resampled skin grayscale image or region image;

[0030] The feature combination layer uses the function pairs in the selected feature combination function FC to extract multiple feature vectors and combine them into one feature vector, which is then output by the output layer.

[0031] Furthermore, the step 3 comprises the following steps:

[0032] Step 3.1: Set the crossover rate to CR and the mutation rate to MR, and set the number of elite GP individuals to M;

[0033] Step 3.2: Input S into the g-th generation population Processed in The fitness value of each GP individual in ;in, express The fitness value of the nth GP individual in;

[0034] Step 3.3: According to The fitness value of each GP individual in is selected The individual with the largest fitness value As the best GP individual of the gth generation and stored in the optimal GP individual set middle;

[0035] Step 3.4: According to ,right Sort all GP individuals in descending order, and select the top individuals in the g-th generation population after sorting. GP individuals as Elite population ;

[0036] Step 3.5: According to and , using the crossover and mutation method to sort the remaining population in the g generation GP individuals evolve, and we get The offspring population ;

[0037] Step 3.6: Input S into the offspring population and follow the procedure in step 3.2 to obtain The fitness value of ,in, express The fitness value of the j-th GP individual in;

[0038] Step 3.7: According to The fitness value of each GP individual in Sort all GP individuals in descending order and select the top GP individuals and elite population The g+1 generation population ;

[0039] Step 3.8: Order Assign to , return to step 3.2 and execute sequentially until the maximum number of iterations is reached until;

[0040] Step 3.9: From The individual corresponding to the maximum fitness value is selected as the global best GP individual.

[0041] Further, the step 3.2 includes the following steps:

[0042] Step 3.2.1: Input S In the nth GP individual, we can get The skin sample image feature vector output by the nth GP individual , thus obtaining Output skin sample image feature vector set ,in, express The output of the kth GP individual The eigenvector of

[0043] Step 3.2.2: Use the max-min normalization method to Perform normalization operation to obtain the normalized feature vector , thus obtaining the normalized feature vector set ;

[0044] Step 3.2.3: Use linear support vector machine SVM to normalize the feature vector set Perform classification and obtain the predicted category label of S;

[0045] Step 3.2.4: Based on the true category label and the predicted category label, use formula (1) to calculate The classification accuracy of the kth GP individual in , and as its fitness value , thus obtaining The fitness value of :

[0046] (1)

[0047] In formula (1), express The feature vector output by the kth GP individual in The total number of correctly classified skin grayscale images.

[0048] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the multi-scale skin abnormality classification method, and the processor is configured to execute the program stored in the memory.

[0049] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the multi-scale skin abnormality classification method when executed by a processor.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention adopts a new flat genetic programming individual structure, including an input layer, a feature extraction layer, a region resampling layer, a feature combination layer and an output layer. The structure proposes a new set of functions and terminal sets to achieve streamlined and effective feature combination in a limited number of feature extractions, thereby assisting relevant personnel in analyzing and identifying important selected features, and realizing the identification and classification of skin abnormality images in real time.

[0052] 2. The regional resampling layer proposed in the present invention performs multi-scale sampling on the image, so that the GP individuals can extract information from global and local features of different scales at the same time, forming a stronger multi-scale feature expression, thereby improving the image classification performance.

[0053] 3. The present invention adopts an evolutionary learning approach to automatically select feature extraction functions in the designed population iteration algorithm to construct a feature extraction combination scheme without the need for domain knowledge and human intervention, thereby reducing the time and cost of image classification and achieving rapid and accurate identification and classification of relevant skin images. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the method of the present invention;

[0055] Figure 2 The GP individual structure diagram is designed for the present invention. DETAILED DESCRIPTION

[0056] In this embodiment, Figure 1 As shown, a multi-scale skin abnormality classification method based on a flattened genetic programming individual structure is performed in the following steps:

[0057] Step 1: Obtain skin image samples with real labels and convert them into grayscale images to construct a skin grayscale image training set ,in, represents the i-th labeled skin grayscale image, express The true category label, express The total number of medium skin grayscale images;

[0058] Obtain skin image samples to be classified and convert them into grayscale images to construct a skin grayscale image test set , represents the jth unlabeled skin grayscale image, represents the total number of skin grayscale images in the test set;

[0059] In this embodiment, the skin image dataset uses PH 2 The dermoscopic image dataset is used as a standard dataset in PH2 There are 200 dermatoscopic images in the dataset, covering three categories of images: common nevus, atypical nevus and malignant melanoma. Each image in the dataset has corresponding category labeling information. 2 70% of the skin image data in the dataset and the corresponding category label information are used to construct the training set S; PH 2 30% of the skin image data in the dataset is used to construct the test set T.

[0060] Step 2: Based on the evolutionary algorithm framework Generate GP individuals and construct the initial population;

[0061] Step 2.1: Build a function set ;

[0062] Function Set Contains: Area resampling function , feature extraction function And feature combination function ;

[0063] Area Resampling Function Includes: Global resampling functions and the local resampling function ;

[0064] Global resampling function Includes: Downsampling function , upsampling function and the clipping function ;in, Used to perform one or two upsampling operations on the skin grayscale image. Used to perform one or two downsampling operations on the skin grayscale image, cropping function Used to adjust the input size of skin grayscale images;

[0065] Local resampling function Includes: Square area resampling function and rectangular area resampling function ;in, It is used to select a square area from the skin grayscale image and perform an upsampling operation. Used to select a rectangular area from the skin grayscale image and perform an upsampling operation;

[0066] Feature extraction function It consists of several image feature extraction functions.

[0067] In this embodiment, the feature extraction function uses five classic image feature extraction operators to describe the texture, edge and other feature information of the image. The input of these functions is the skin grayscale image or the resampled area, and the output is the feature vector. Specifically, they include:

[0068] 1) HOG (Histogram of Oriented Gradients): HOG is a classic image feature extraction operator. Its basic principle is to describe the texture and edge information of the image by calculating the gradient direction histogram of the local area of ​​the image. In terms of specific parameter settings, the default configuration is used: the number of feature directions is set to 9, the pixel size of each cell unit is 8×8, and the number of cell units contained in each block is set to 3×3, and the "L2-Hys" normalization method is used.

[0069] 2)uLBP (Uniform Local Binary Patterns): uLBP introduces the concept of uniform patterns based on LBP (local binary patterns) to enhance the feature discrimination ability and anti-noise performance.

[0070] 3) Gabor function: It includes two parts, Gabor_real and Gabor_imag, which can extract scale-invariant and direction-selective features from images from multiple angles. Such extracted features can provide a more comprehensive and robust representation, which helps improve classification performance.

[0071] 4) SIFT (Scale Invariant Feature Transform): The SIFT function has the main characteristics of scale invariance and rotation invariance, and can effectively extract local features with good discriminability and stability.

[0072] 5) Hist (Histogram): The Hist function is used to calculate the distribution of pixel values ​​in an image. This function divides the pixel value range of an image into multiple discrete intervals and calculates the number or proportion of pixels in each interval to form a histogram. Histograms can be used to describe image characteristics such as brightness, contrast, and color distribution.

[0073] The feature combination function FC is composed of several image feature combination functions, which is used to combine the feature extraction function The extracted features are combined into a feature vector;

[0074] In this embodiment, two feature combination functions Fea_Con4 and Fea_Con5 are used, which can respectively receive 4 or 5 multi-scale features of the feature extraction layer and combine them into a feature vector. The reason for limiting the connection to 4 or 5 features is that the area resampling function may obtain 4 or 5 input images of different scales, thereby generating a corresponding number of multi-scale features.

[0075] Step 2.2: Build Includes: skin grayscale image, horizontal and vertical coordinate parameters of the upper left corner rectangular area of ​​R_Up_Sample in LRS, horizontal and vertical coordinate parameters of the upper left corner of the square area in S_Up_Sample, and the size of the rectangular area , Long parameters for adjusting skin grayscale images in and width parameters ;

[0076] In this embodiment, set and The horizontal and vertical coordinate parameters of the upper left corner rectangular area are in the range of [0, boundx - 10] and [0, boundy - 10], where boundx and boundy represent the width and length of the skin grayscale image respectively; The value range of is smaller than the size of the skin grayscale image; Long parameter and width parameters The parameter values ​​of are in the range of [0, boundx] and [0, boundy]. These parameters randomly select specific values ​​during the evolution process and participate in the evolution of GP individuals.

[0077] Step 2.2: Set the maximum number of iterations , assign 50 to , the current number of iterations is ,initialization ;

[0078] Step 2.3: Based on the function set Funcs and the terminal set Terminals, use Algorithm generation Individual trees as GP individuals, assign 100 to , thus forming the g-th generation population ;

[0079] like Figure 2 As shown, the structure of constructing any individual tree includes: input layer, region resampling layer, feature extraction layer, feature combination layer and output layer in sequence;

[0080] The generation method of any individual tree is as follows: select several functions from the regional resampling function RS as the nodes of the regional resampling layer in the individual tree; select several functions from the feature extraction function FE as the nodes of the feature extraction layer in the individual tree, and connect them with the nodes of the regional resampling layer; select a function from the feature combination function FC as the node of the feature combination layer in the individual tree, and connect it with the nodes of the feature extraction layer; according to the terminal set Define the random parameters of the input layer and the regional resampling layer in the individual tree; the above selection process is randomly selected by GP individuals during the evolution process, thereby obtaining a GP individual tree; wherein,

[0081] The input layer is used to input the skin grayscale image;

[0082] The regional resampling layer uses the selected regional resampling function The function in resamples the skin grayscale image or the regional image in the skin grayscale image input by the input layer;

[0083] The feature extraction layer uses the selected feature extraction function The function in extracts features from the resampled skin grayscale image or region image;

[0084] The feature combination layer uses the function pairs in the selected feature combination function FC to extract multiple feature vectors and combine them into one feature vector, which is then output by the output layer.

[0085] Step 3: Evolve the g-th generation population using evolutionary learning methods GP individuals among them, find the best GP individual;

[0086] Step 3.1: Set the crossover rate to CR and the mutation rate to MR, assign 0.8 to CR, assign 0.19 to MR, set the number of elite GP individuals to M, and assign M to 10.

[0087] Step 3.2: Input S into the g-th generation population Processed in The fitness value of each GP individual in ;in, express The fitness value of the nth GP individual in;

[0088] Step 3.2.1: Input S In the nth GP individual, we can get The skin sample image feature vector output by the nth GP individual ; thus obtaining Output skin sample image feature vector set ,in, express The output of the kth GP individual The eigenvector of

[0089] Step 3.2.2: Use the max-min normalization method to Perform normalization operation to obtain the normalized feature vector ; Thus, we get the normalized feature vector set ;

[0090] Step 3.2.3: Use linear support vector machine SVM to normalize the feature vector set Perform classification and obtain the predicted category label of S.

[0091] Step 3.2.4: Based on the true category label and the predicted category label, use formula (1) to calculate The classification accuracy of the kth GP individual in , and as its fitness value , thus obtaining The fitness value of :

[0092]

[0093] In formula (1), express The feature vector output by the kth GP individual in The total number of correctly classified skin grayscale images.

[0094] Step 3.3: According to The fitness value of each GP individual in ,choose The individual with the largest fitness value As the best GP individual of the gth generation, it is stored in the optimal GP individual set middle;

[0095] Step 3.4: According to right Sort all GP individuals in descending order, and select the top individuals in the g-th generation population after sorting. GP individuals as Elite population ;

[0096] Step 3.5: According to and , using the crossover and mutation method to sort the remaining population in the g generation GP individuals evolve, and we get The offspring population .

[0097] Step 3.6: Input S into the offspring population and follow the procedure in step 3.2 to obtain The fitness value of ,in, express The fitness value of the j-th GP individual in;

[0098] Step 3.7: According to The fitness value of each GP individual in Sort all GP individuals in descending order and select the top GP individuals and elite population The g+1 generation population ;

[0099] Step 3.8: Order Assign to , return to step 3.2 and execute sequentially until the maximum number of iterations is reached until;

[0100] Step 3.9: From The individual corresponding to the maximum fitness value is selected as the global best GP individual.

[0101] Step 4: The individual tree structure corresponding to the global best GP individual is used as the optimal skin image classification model, and is used to process the test set T to obtain the feature vector of T After that, the linear SVM classifier is input for classification, and we get The predicted classification label of Indicates T The feature vector of .

[0102] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the 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 stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0104] According to the classification results of the test set, combined with the test set in PH 2 The real labels in the data set are obtained by the method of the present invention in PH 2 The classification accuracy on the dermoscopic image dataset is shown in Table 1. The classification accuracy of the proposed method is better than that of the traditional classification methods (NB / SVM / LMLP), and is also better than the GP-based COGP and FLGP methods;

[0105] Table 1 Classification accuracy of various classification methods on the PH2 dermoscopy image dataset

[0106] Skin image abnormality classification method Method of the present invention NB SVM MLP COGP FLGP Classification Accuracy 0.8802 0.7719 0.6413 0.7893 0.6712 0.7814

[0107] It can be seen from Table 1 that the classification accuracy of the method of the present invention is greatly improved compared with other methods. The proposed multi-scale feature skin abnormality classification method can select features of different scales based on the GP program structure of the present invention and combine them into more discriminating high-level features, thereby improving the classification performance of skin abnormalities and assisting people to identify and classify skin abnormalities without the need for domain knowledge.

Claims

1. A multi-scale skin abnormality classification method based on a flattened genetic programming individual structure, characterized in that: The steps include: Step 1: Obtain skin image samples with real labels and convert them into grayscale images to construct a skin grayscale image training set ,in, represents the i-th labeled skin grayscale image, express The true category label, express The total number of medium skin grayscale images; Obtain skin image samples to be classified and convert them into grayscale images to construct a skin grayscale image test set , represents the jth unlabeled skin grayscale image, represents the total number of skin grayscale images in the test set; Step 2: Based on the evolutionary algorithm framework Generate GP individuals and construct the initial population; Step 3: Evolve the g-th generation population using evolutionary learning methods GP individuals in, find the global best GP individual; Step 4: The individual tree structure corresponding to the global best GP individual is used as the optimal skin image classification model, and is used to process the test set T to obtain the feature vector of T After that, the linear SVM classifier is input for classification, and we get The predicted classification label of Indicates T The feature vector of .

2. A multi-scale skin abnormality classification method based on flattened genetic programming individual structure according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2.1: Build a function set ; The function set Contains: Area resampling function , feature extraction function and feature combination function ; The area resampling function Includes: Global resampling functions and the local resampling function ; The global resampling function Includes: Downsampling function , upsampling function and the clipping function ;in, Used to perform one or two upsampling operations on the skin grayscale image. Used to perform one or two downsampling operations on the skin grayscale image, cropping function Used to adjust the input size of skin grayscale images; The local resampling function Includes: Square area resampling function and rectangular area resampling function ;in, It is used to select a square area from the skin grayscale image and perform an upsampling operation. Used to select a rectangular area from the skin grayscale image and perform an upsampling operation; The feature extraction function It is composed of several image feature extraction functions; The feature combination function FC is composed of several image feature combination functions, which is used to combine the feature extraction function The extracted features are combined into a feature vector; Step 2.2: Construct a terminal set Terminals to define the parameters of related functions in the function set; the terminal set Includes: skin grayscale image, horizontal and vertical coordinate parameters of the upper left corner rectangular area of ​​R_Up_Sample in LRS, horizontal and vertical coordinate parameters of the upper left corner of the square area in S_Up_Sample, and the size of the rectangular area , Long parameters for adjusting skin grayscale images in and width parameters ; Step 2.2: Set the maximum number of iterations , the current number of iterations is ,initialization ; Step 2.3: Based on the function set Funcs and the terminal set Terminals, use Algorithm generation Individual trees as GP individuals, thus forming the g-th generation population ; The structure of constructing any individual tree includes: input layer, region resampling layer, feature extraction layer, feature combination layer and output layer; The generation method of any individual tree is as follows: select several functions from the regional resampling function RS as nodes of the regional resampling layer in the individual tree; select several functions from the feature extraction function FE as nodes of the feature extraction layer in the individual tree, and connect them with the nodes of the regional resampling layer; select a function from the feature combination function FC as a node of the feature combination layer in the individual tree, and connect it with the nodes of the feature extraction layer; according to the terminal set Define the parameters of the input layer and regional resampling layer in the individual tree, thereby obtaining an individual tree; wherein: The input layer is used to input a skin grayscale image; The regional resampling layer uses the selected regional resampling function The function in resamples the skin grayscale image or the regional image in the skin grayscale image input by the input layer; The feature extraction layer uses the selected feature extraction function The function in extracts features from the resampled skin grayscale image or region image; The feature combination layer uses the function pairs in the selected feature combination function FC to extract multiple feature vectors and combine them into one feature vector, which is then output by the output layer.

3. A multi-scale skin abnormality classification method based on flattened genetic programming individual structure according to claim 2, characterized in that: The step 3 comprises the following steps: Step 3.1: Set the crossover rate to CR and the mutation rate to MR, and set the number of elite GP individuals to M; Step 3.2: Input S into the g-th generation population Processed in The fitness value of each GP individual in ;in, express The fitness value of the nth GP individual in ; Step 3.3: According to The fitness value of each GP individual in is selected The individual with the largest fitness value As the best GP individual of the gth generation and stored in the optimal GP individual set middle; Step 3.4: According to ,right Sort all GP individuals in descending order, and select the top individuals in the g-th generation population after sorting. GP individuals as Elite population ; Step 3.5: According to and , using the crossover and mutation method to sort the remaining population in the g generation GP individuals evolve, and we get The offspring population ; Step 3.6: Input S into the offspring population and follow the procedure in step 3.2 to obtain The fitness value of ,in, express The fitness value of the j-th GP individual in; Step 3.7: According to The fitness value of each GP individual in Sort all GP individuals in descending order and select the top GP individuals and elite population The g+1 generation population ; Step 3.8: Order Assign to , return to step 3.2 and execute sequentially until the maximum number of iterations is reached until; Step 3.9: From The individual corresponding to the maximum fitness value is selected as the global best GP individual.

4. A multi-scale skin abnormality classification method based on flattened genetic programming individual structure according to claim 3, characterized in that: The step 3.2 comprises the following steps: Step 3.2.1: Input S In the nth GP individual, we can get The skin sample image feature vector output by the nth GP individual , thus obtaining Output skin sample image feature vector set ,in, express The output of the kth GP individual The eigenvector of Step 3.2.2: Use the max-min normalization method to Perform normalization operation to obtain the normalized feature vector , thus obtaining the normalized feature vector set ; Step 3.2.3: Use linear support vector machine SVM to normalize the feature vector set Perform classification and obtain the predicted category label of S; Step 3.2.4: Based on the true category label and the predicted category label, use formula (1) to calculate The classification accuracy of the kth GP individual in , and as its fitness value , thus obtaining The fitness value of : (1) In formula (1), express The feature vector output by the kth GP individual in The total number of correctly classified skin grayscale images.

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the multi-scale skin abnormality classification method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-scale skin abnormality classification method according to any one of claims 1 to 4 are performed.

Images