CN116258697B - Device and method for automatic classification of children's skin disease images based on coarse annotation - Google Patents

Abstract

The invention discloses an automatic classification device and method for children's skin disease images based on coarse annotation. After coarse annotation of the lesion area of the acquired child skin disease image, the coarse annotation of the lesion area is preprocessed to establish a mask mask. Label images; build a classification model including U‑Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module, and feature fusion and classification module, using mask masks Annotated images are used for supervised learning of the classification model to optimize the parameters of the classification model; the parameter-optimized classification model is used to automatically classify children's skin disease images. The device and method build a model that can accurately and automatically classify children's skin diseases based on coarse annotations, and improve the classification accuracy of children's skin disease images.

Description

Device and method for automatic classification of children's skin disease images based on coarse annotation

Technical field

The invention belongs to the technical field of image classification, and specifically relates to an automatic classification device and method for children's skin disease images based on coarse annotation.

Background technique

Skin diseases are among the most common diseases in children. At present, the skin diseases with the highest incidence rate and the most common among children's skin diseases include: atopic dermatitis (AD), urticaria (Urticaria), hemangioma, diaper dermatitis (Diaper Dermatitis), etc. These four diseases seriously affect the quality of life of children, harm their physical and mental health, and even endanger their lives, causing heavy economic and social burdens.

At present, most children with skin diseases are first diagnosed by non-dermatologists at the grassroots level. However, limited by the experience and ability of grassroots doctors and the incomplete medical facilities of hospitals, the misdiagnosis rate of children's skin diseases at the grassroots level is high. It is difficult to solve this problem through traditional clinical methods in the short term. Therefore, exploring new methods to accurately and quickly diagnose children's skin diseases and quickly improve the accuracy of first diagnosis by grassroots doctors has become an important issue that needs to be solved urgently.

There are significant differences between children's skin and adult skin, mainly including: 1. Skin structure and barrier differences: the average thickness of children's stratum corneum (about 7 μm) is 30% thinner than that of adults (about 10 μm), and the epidermis is 20% thinner. With small-volume keratinocytes and thinner stratum corneum, the skin barrier function is immature, the water loss rate is fast, the resistance is weak, and the reaction to external stimuli is strong; 2. Differences in skin composition: The concentration of natural moisturizing factors in children's skin keratinocytes is obvious It is lower than adults and has lower sebum levels. The total lipid content and the amount of sebum secreted are less than those of adults, so children's skin is more likely to be dry and dehydrated. At the same time, the melanin content in children's skin is lower than that of adults, making them more susceptible to sun damage; 3. Weak microbial barrier: the pH value of children's skin is close to neutral (6.6-7.7); the microbiome is not yet stable and easily affected. Based on the above characteristics, more exploration of children's skin images can be conducted.

Currently, skin disease image recognition models based on deep learning have been widely researched and tested, but the recognition capabilities are limited, and most research methods are still focused on dealing with adult skin disease problems and do not involve children's skin problems. In addition, research methods based on deep learning require a large amount of manpower for detailed annotation, which places high requirements on the accuracy of training data. However, skin disease problems have damaging or discontinuous annotation characteristics, so detailed annotation is required, which is costly. A lot of manpower and material resources. In addition, there are still situations where the edges are not clear, and artificial annotation may lead to imperfect recognition. Therefore, it is particularly important to propose an automatic classification method for children's skin disease images that can achieve fine annotation using rough annotation.

Contents of the invention

In view of the above, the purpose of the present invention is to provide an automatic classification device and method for children's skin disease images based on rough annotation, build a model that can accurately and automatically classify children's skin diseases based on rough annotation, and improve the classification accuracy of children's skin disease images. .

In order to achieve the above-mentioned object of the invention, the automatic classification method of children's skin disease images based on coarse annotation provided by the embodiment includes the following steps:

After coarse annotation of the lesion area on the acquired children's skin disease images, preprocess the coarse annotation of the lesion area to establish a mask annotation image;

Construct a classification model including U-Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module, and feature fusion and classification module. Among them, U-Net, texture feature The extraction module, color feature extraction module, and shape color feature extraction module are respectively used to extract depth features, texture features, color features, and shape features from the mask annotation image. The first correlation analysis module is used to extract texture features and color features. Perform primary feature fusion, the second correlation analysis module is used to perform secondary feature fusion on the shape feature and the primary feature fusion result, and the feature fusion and classification module is used to perform feature fusion on the splicing result of the secondary feature fusion result and the depth feature. perform classification of childhood dermatoses;

Use mask annotation images to perform supervised learning on the classification model to optimize the parameters of the classification model;

Automatic classification of children's skin disease images using a parameter-optimized classification model.

In an optional embodiment, preprocessing the coarsely annotated lesion area to create a mask annotation image includes:

Using the channel differences between children's skin and adult skin on RGB images, the CLAHE algorithm is used to process children's skin disease images to obtain preprocessed images;

After using the SLIC algorithm to perform superpixel segmentation on the roughly labeled lesion area in the preprocessed image, the three-channel values of each pixel in the superpixel block are added to obtain the single-pixel chromatographic value;

The clustering algorithm is used to divide the chromatographic value of a single pixel into two categories, that is, the pixels in the lesion area are divided into 0 or 1 pixels to obtain a binary image;

Search all boundaries based on binary images, calculate the area within each boundary, and then fill the maximum connected domain to obtain a fine-grained mask annotation image.

In an optional embodiment, the clustering algorithm includes a K-means clustering algorithm, that is, the K-means clustering algorithm is used to divide the chromatographic value of a single pixel into two categories.

In an optional embodiment, in the texture feature extraction module, the color histogram feature extraction method based on Bayesian classifier is used to extract color features, including:

Based on the HIS color model value of each pixel in the mask mask annotation image, three vectors X ^HS , X ^HI and X ^SI are obtained, and the three vectors are classified into 9 color category histograms through the Bayesian classifier , count the number of pixels in the histogram of each color category to obtain color features.

In an optional embodiment, in the texture feature extraction module, the ULBP algorithm is used to extract texture features of the mask annotation image.

In an optional embodiment, in the shape feature extraction module, shape features based on regional edges are extracted from the mask annotation image, including shape parameters, bending energy, rectangularity, and circularity.

In an optional embodiment, the first correlation analysis module and the second correlation analysis module use a CCA algorithm to perform feature fusion.

In an optional embodiment, the feature fusion and classification module includes at least one convolutional layer and a fully connected layer. Feature fusion is implemented through the convolutional layer, and skin disease classification is implemented through the fully connected layer.

In an optional embodiment, when using mask annotation images to perform supervised learning on a classification model, the cross-entropy of the children's skin disease classification results output by the classification model and the true classification label is used as the loss function to update the model parameters.

In order to achieve the above object of the invention, the embodiment provides an automatic classification device for children's skin disease images based on coarse annotation, including a preprocessing unit, a model building unit, a parameter optimization unit and an application unit;

The preprocessing unit is used to perform rough annotation of the lesion area on the acquired children's skin disease images, and then preprocess the roughly annotated lesion area to establish a mask annotation image;

The model building unit is used to build a classification model including U-Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module and feature fusion and classification module, where , U-Net, texture feature extraction module, color feature extraction module, and shape color feature extraction module are respectively used to extract depth features, texture features, color features, and shape features from the mask mask annotation image. The first correlation analysis module uses The second correlation analysis module is used for primary feature fusion of texture features and color features. The second correlation analysis module is used for secondary feature fusion of shape features and primary feature fusion results. The feature fusion and classification module is used for secondary feature fusion results and depth features. The splicing results are used for feature fusion before classifying children's skin diseases;

The parameter optimization unit is used to perform supervised learning on the classification model using mask annotation images to optimize the parameters of the classification model;

The application unit is used to automatically classify children's skin disease images using a parameter-optimized classification model.

Compared with the prior art, the beneficial effects of the present invention include at least:

By preprocessing the coarsely annotated lesion area to establish a fine-grained mask annotation image, only coarse annotation of the larger lesion area can effectively allow the classification model to obtain sufficient semantic information without the need for a large amount of The precise annotation can achieve accurate automatic classification of children's skin disease images, solving the problem of difficult precise annotation of skin diseases due to unclear boundaries. Furthermore, by combining texture features, color features and shape features in the classification model to classify children's skin diseases, it can improve the classification accuracy of children's skin disease images and provide doctors with the ability to assist in diagnosis.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of the automatic classification method of children's skin disease images based on coarse annotation provided by the embodiment;

Figure 2 is a schematic structural diagram of the classification model provided by the embodiment;

Figure 3 is an example diagram of a color channel histogram of children's skin diseases provided by the embodiment;

Figure 4 is a comparison chart of various preprocessing methods for children's skin disease images provided by the embodiment;

Figure 5 is an example diagram of super-pixel segmentation of coarse annotation areas of children's skin disease images provided by the embodiment;

Figure 6 is a schematic structural diagram of an automatic classification device for children's skin disease images based on coarse annotation provided by the embodiment.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present 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 here are only used to explain the present invention and do not limit the scope of the present invention.

In order to overcome the shortcomings in the existing technology that require a large number of accurate children's skin disease annotations, which results in great cost, and at the same time, in order to achieve accurate classification of children's skin disease images. The embodiment provides a method and device for automatically classifying children's skin disease images based on rough annotation. The image feature factors of children's skin that are different from adult skin are used for preliminary preprocessing, and then the graphics method is used to compare the rough annotation with the actual image pixels. Match, obtain semantic segmentation images with high accuracy, and establish mask annotation images. Taking advantage of the image features of skin disease images, which are mostly flaky and dot-like, a deep learning training method based on image feature fusion is used, using Unet as the backbone network, and then using traditional image processing methods to label the color features and texture of the image on the mask. Features and shape features are subjected to canonical correlation analysis (CCA) to obtain image fusion features and added to the deep learning network to establish a complete classification network. Using mask annotation images constructed based on coarse annotation to train the classification model can achieve higher training accuracy, which can greatly save doctors' time and energy for fine annotation while ensuring the effective recognition ability of the classification model.

As shown in Figure 1, the automatic classification method of children's skin disease images based on coarse annotation provided by the embodiment includes the following steps:

S110. After rough annotation of the lesion area on the acquired children's skin disease image, preprocess the rough annotation of the lesion area to establish a mask annotation image.

Coarse annotation means that the doctor drags and selects a rectangular frame to annotate a larger lesion area. The lesion area selected by this rough annotation is relatively rough and may include non-lesion areas, or the lesion area may not be comprehensively selected. After the rough annotation of the lesion area, preprocess the rough annotation of the lesion area to establish a mask annotation image, including:

(a) Using the channel difference between children's skin and adult skin on the RGB image, using the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm on the channel to process children's skin disease images to obtain preprocessing image.

Through experimental research, it was found that because children's skin has less melanin, the value of the RGB spectrum of normal skin pictures on the G channel will be significantly higher than the values on the other two channels, as shown in Figure 3, which will cause subsequent ultrasonic During pixel segmentation, the edge information of the lesion is lost due to the large difference from the sum of RGB three-channel pixels in the lesion area. In addition, because skin disease images are generally taken by ordinary cameras, the scene and light changes are greatly different. Using the CLAHE algorithm can increase the contrast of subsequent lesion area processing, and can also limit the contrast amplification in the image, thereby reducing noise. Zooming in on the problem, the comparison of operations is shown in Figure 4.

(b) After using the SLIC (simple linear iterative clustering) algorithm to perform superpixel segmentation on the coarsely annotated lesion area in the preprocessed image (for example, dividing the lesion area into superpixel blocks of less than 200 units), based on the coarsely annotated annotation data, The three-channel values of each pixel in the superpixel block are added to obtain the single-pixel chromatographic value.

The present invention adopts a superpixel segmentation method to refine the coarsely labeled images. A superpixel refers to a small area composed of a series of adjacent pixels with similar characteristics such as color, brightness, texture, etc. Most of these small areas retain effective information for further image segmentation, and generally do not destroy the boundary information of objects in the image. A small number of superpixels are used instead of a large number of pixels to express image features, which reduces the complexity of image processing. The superpixel blocks generated by the superpixel segmentation SLIC algorithm are relatively compact, and domain features are easy to express. At the same time, it requires few parameters to be set and adjusted, is simple to operate, and is fast. It has a good effect on image compactness and contour maintenance. In addition, SLIC The algorithm is compatible with segmentation of grayscale images and color images. Generally speaking, the superpixel segmentation method is a means of blurring using fine grain images, but this invention attempts to use its principle to perform a reverse refinement operation, trying to treat the image within the coarse annotation frame as a whole block of pixels. The super-pixel segmentation method is used to perform pixel segmentation with a high number of units to achieve fine segmentation of rough boundary annotations, laying the foundation for fine-grained implementation of converting coarse annotations into fine annotations. The superpixel segmentation results of the coarsely labeled area can be shown in Figure 5.

(c) Use clustering algorithms such as K-means to divide the chromatographic value of a single pixel into two categories, that is, divide the pixels in the lesion area into 0 or 1 pixels to obtain a binary image.

(d) Search all boundaries based on binary images, calculate the area within each boundary, and then fill in the maximum connected domain to obtain a fine-grained mask annotation image.

In order to better process finely labeled boundary information, the present invention further processes coarsely labeled images by using a clustering algorithm and a method of finding connected domains. By using the K-means clustering algorithm, the chromatographic value of a single pixel is obtained by adding the three-channel values of each pixel, and the superpixel blocks obtained by superpixel segmentation are clustered into two major categories, that is, within the rough label box The data is divided into 0/1 pixels. Due to problems such as image peaks and critical conditions, the binarized rough annotation data will have discontinuous distribution. Therefore, the present invention further uses the method of finding the maximum connected domain, using the findContours function to find all boundaries in the binary image, then calculating the area within each boundary through the contourArea function, and finally filling the maximum connected domain through the fillConvexPoly function. Through the above steps, the goal of automatically refining the lesion boundary image based on the coarse annotation image is achieved.

S120, build a classification model.

In the embodiment, the classification model is designed based on the U-Net network. In order to correct the errors caused by the refinement of rough annotations, an image feature fusion method is added based on the U-Net network, that is, image features are used as one of the training parameters, and an overall classification model as shown in Figure 2 is established.

As shown in Figure 2, the classification model includes U-Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module, and feature fusion and classification module. Among them, U-Net, texture feature extraction module, color feature extraction module, and shape color feature extraction module are respectively used to extract depth features, texture features, color features, and shape features from mask mask annotation images. The first correlation analysis module It is used for primary feature fusion of texture features and color features. The second correlation analysis module is used for secondary feature fusion of shape features and primary feature fusion results. The feature fusion and classification module is used for secondary feature fusion results and depth features. The splicing results are feature fused and then classified into children's skin diseases.

In the texture feature extraction module, the color histogram feature extraction method based on Bayesian classifier is used to extract color features, including:

Scan each pixel in the mask annotated image, and obtain three vectors X ^HS , X ^HI and X ^SI based on the HIS color model value (h, s, i) of each pixel in the mask annotated image. , where, X ^HS = [h ^{, s], X HI =} [h, i], , that is, classify the three vectors into 9 color category histograms according to the Bayesian classifier, and count the number of pixels in each color category histogram to obtain the color features. Among them, the 9 color categories include white, red, light brown, dark brown, light blue gray, dark blue gray, black and undefined color.

In the texture feature extraction module, the ULBP (Uniform Local Binary Pattern) algorithm is used to extract the texture features of the mask annotation image. The ULBP algorithm is defined as:

Among them, g _c is the gray value of the central pixel, R is the radius of the area, g _i is the gray value of the i-th pixel in the area of g _c , and P is the number of pixels in the neighborhood of the central pixel, uniform The operator U is used to represent the number of conversions between 0 and 1 in the local binary pattern. The superscript riu2 reflects that the rotationally invariant "uniform" mode has a U value of at most 2.

The ULBP algorithm is an algorithm that measures the relationship between pixels and neighboring pixels. It extracts texture features by analyzing the relationship between pixels and neighboring pixels. The distribution of these local textures can form the overall texture of the image, and finally the texture feature vector of the image can be obtained by counting each neighborhood structure.

In the shape feature extraction module, shape features based on regional edges are extracted from the mask annotation image, including shape parameters, bending energy, rectangularity, and circularity.

Among them, the expression of the shape parameter F is:

Among them, C is the perimeter of the region boundary, and A is the area of the region.

The expression of bending energy B is:

Among them, p is the arc length parameter, P is the total length of the curve, and k(p) is the curvature function.

Rectangularity R is the ratio of the area of the region to the area of the smallest circumscribed rectangle, indicating the degree of saturation of the circumscribed rectangle. Its expression is:

Among them, A is the area of the region, and A _box is the area of the smallest circumscribed rectangle.

Circularity C is used to measure the circularity of an area, and its expression is:

Among them, P is the perimeter of the region boundary, and A is the area of the region.

Through the above method, image features such as color features, texture features, and shape features are extracted. In order to unify the feature dimensions, the canonical correlation analysis method CCA (Canonical Correlation Analysis) is used to perform feature fusion in the first correlation analysis module and the second correlation analysis module. CCA is a statistical analysis method used to analyze the relationship between two random vectors. It can remove redundant information between features and fuse features well.

The feature fusion and classification module is connected to the output of the second correlation analysis module and U-Net, including at least one convolutional layer and a fully connected layer, such as 3 convolutional layers and 1 fully connected layer, implemented through the convolutional layer Feature fusion realizes skin disease classification through fully connected layers.

S130: Use the mask annotation image to perform supervised learning on the classification model to optimize the parameters of the classification model.

In the embodiment, when the mask annotated image is used to perform supervised learning on the classification model, the cross-entropy function of the children's skin disease classification results output by the classification model and the classification true value label is used as a loss function to update the model parameters. The model parameters include After parameter optimization of U-Net's network parameters, feature fusion and classification modules, a reliable classification model was obtained, which can accurately classify children's skin disease images.

S140, use the parameter-optimized classification model to automatically classify children's skin disease images.

In the embodiment, when a parameter-optimized classification model is used to automatically classify children's skin disease images, the obtained children's skin disease images to be classified are subjected to rough annotation of the lesion area, and then the rough annotation of the lesion area is preprocessed to establish a mask. The mask annotated image is input into the parameter-optimized classification model, and the classification results of children's skin disease images are obtained through forward inference calculation.

Based on the same inventive concept, the embodiment also provides an automatic classification device for children's skin disease images based on coarse annotation, as shown in Figure 6, including a preprocessing unit, a model building unit, a parameter optimization unit and an application unit; wherein, the preprocessing unit The processing unit is used to perform rough annotation of the lesion area on the acquired children's skin disease images, and then pre-processes the rough annotation of the lesion area to establish a mask annotation image; the model construction unit is used to build a classification model; the parameter optimization unit is used to It uses mask mask annotation images to perform supervised learning on the classification model to optimize the parameters of the classification model; the application unit is used to use the parameter-optimized classification model to automatically classify children's skin disease images.

It should be noted that when the automatic classification device for children's skin disease images provided in the above embodiments performs automatic classification of children's skin disease images, the above-mentioned division of each functional unit should be used as an example. The above functions can be allocated to different functions as needed. Unit completion means that the internal structure of the terminal or server is divided into different functional units to complete all or part of the functions described above. In addition, the device for automatically classifying children's skin disease images provided in the above embodiments and the embodiment of the method for automatically classifying children's skin disease images belong to the same concept. For details of the implementation process, see the embodiment of the method for automatically classifying children's skin disease images, which will not be described again here.

The above-described specific embodiments describe in detail the technical solutions and beneficial effects of the present invention. It should be understood that the above are only the most preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, equivalent substitutions, etc. made within the scope of the invention shall be included in the protection scope of the present invention.

Claims (5)

1. An automatic classification method for children's skin disease images based on coarse annotation, which is characterized by including the following steps: After rough annotation of the lesion area on the acquired children's skin disease images, pre-process the rough annotation of the lesion area to establish a mask annotation image, including: using the channel difference between children's skin and adult skin on the RGB image, The CLAHE algorithm is used to process children's skin disease images to obtain a preprocessed image; the SLIC algorithm is used to perform superpixel segmentation on the roughly labeled lesion area in the preprocessed image, and then the three-channel values of each pixel in the superpixel block are added to obtain Single pixel chromatographic value; use a clustering algorithm to divide the single pixel chromatographic value into two categories, that is, divide the pixels in the lesion area into 0 or 1 pixels to obtain a binary image; search all boundaries based on the binary image, and Calculate the area within each boundary and then fill the maximum connected domain to obtain a fine-grained mask annotation image; Construct a classification model including U-Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module, and feature fusion and classification module. Among them, U-Net, texture feature The extraction module, color feature extraction module, and shape color feature extraction module are respectively used to extract depth features, texture features, color features, and shape features from the mask annotation image. The first correlation analysis module is used to extract texture features and color features. Perform primary feature fusion, the second correlation analysis module is used to perform secondary feature fusion on the shape feature and the primary feature fusion result, and the feature fusion and classification module is used to perform feature fusion on the splicing result of the secondary feature fusion result and the depth feature. Classify children's skin diseases; wherein, in the texture feature extraction module, a color histogram feature extraction method based on a Bayesian classifier is used to extract color features, including: labeling each pixel in the image based on a mask mask The HIS color model value obtains three vectors X ^HS , X ^{HI and} , to obtain color features; in the texture feature extraction module, the ULBP algorithm is used to extract the texture features of the mask mask annotation image; in the shape feature extraction module, the mask mask annotation image is extracted based on regional edges Shape features include shape parameters, bending energy, rectangularity, and circularity; the first correlation analysis module and the second correlation analysis module use the CCA algorithm to perform feature fusion; Use mask annotation images to perform supervised learning on the classification model to optimize the parameters of the classification model; Automatic classification of children's skin disease images using a parameter-optimized classification model. 2. The automatic classification method of children's skin disease images based on rough annotation according to claim 1, characterized in that the clustering algorithm includes a K-means clustering algorithm, that is, the K-means clustering algorithm is used to classify single pixels. Chromatographic values are divided into two categories. 3. The automatic classification method of children's skin disease images based on coarse annotation according to claim 1, characterized in that the feature fusion and classification module includes at least one convolution layer and a fully connected layer, and feature fusion is achieved through the convolution layer. , achieving skin disease classification through fully connected layers. 4. The automatic classification method of children's skin disease images based on coarse annotation according to claim 1, characterized in that when using mask mask annotation images to perform supervised learning on the classification model, the children's skin disease classification results output by the classification model are and the cross-entropy of the classification ground truth labels are used as the loss function to update the model parameters. 5. An automatic classification device for children's skin disease images based on coarse annotation, characterized by including a preprocessing unit, a model building unit, a parameter optimization unit and an application unit; The preprocessing unit is used to perform rough annotation of the lesion area on the acquired children's skin disease images, and then preprocess the roughly annotated lesion area to establish a mask annotation image, including: using the method based on children's skin and adult skin. For channel differences in RGB images, the CLAHE algorithm is used to process children's skin disease images to obtain preprocessed images; after using the SLIC algorithm to perform superpixel segmentation on the roughly labeled lesion areas in the preprocessed image, each pixel in the superpixel block is The three-channel values are added to obtain the single-pixel chromatographic value; the clustering algorithm is used to divide the single-pixel chromatographic value into two categories, that is, the pixels in the lesion area are divided into 0 or 1 pixels to obtain a binary image; based on the binary The value image searches all boundaries, calculates the area within each boundary, and then fills the maximum connected domain to obtain a fine-grained mask annotation image; The model building unit is used to build a classification model including U-Net, texture feature extraction module, color feature extraction module, shape feature extraction module, first correlation analysis module, second correlation analysis module and feature fusion and classification module, where , U-Net, texture feature extraction module, color feature extraction module, and shape color feature extraction module are respectively used to extract depth features, texture features, color features, and shape features from the mask mask annotation image. The first correlation analysis module uses The second correlation analysis module is used for primary feature fusion of texture features and color features. The second correlation analysis module is used for secondary feature fusion of shape features and primary feature fusion results. The feature fusion and classification module is used for secondary feature fusion results and depth features. After feature fusion of the splicing results, children's skin diseases are classified; wherein, in the texture feature extraction module, the color histogram feature extraction method based on Bayesian classifier is used to extract color features, including: based on mask mask annotation The HIS color model value of each pixel in the image is obtained as three vectors: X ^HS , X ^{HI and} The number of pixels in the histogram is used to obtain color features; in the texture feature extraction module, the ULBP algorithm is used to extract the texture features of the mask mask annotated image; in the shape feature extraction module, the mask mask is The annotated image extracts shape features based on regional edges, including shape parameters, bending energy, rectangularity, and circularity; the first correlation analysis module and the second correlation analysis module use the CCA algorithm to perform feature fusion; The parameter optimization unit is used to perform supervised learning on the classification model using mask annotation images to optimize the parameters of the classification model; The application unit is used to automatically classify children's skin disease images using a parameter-optimized classification model.