Acne detection method and system based on agency dynamic mixing normalization

Abstract

The invention discloses an agent dynamic hybrid normalization-based acne detection method and system, and relates to the technical field of image processing, and the method comprises the steps: dividing a data set through hierarchical random sampling, carrying out the data enhancement, and constructing a YOLOv11-based acne detection network model which comprises a backbone network, an agent dynamic hybrid normalization module and a Neck module, the backbone network extracts multi-scale features, the proxy dynamic hybrid normalization module combines instance normalization and batch normalization dynamic weighted fusion to realize feature optimization, the Neck module performs cross-scale fusion on the normalized features to improve spatial expression and semantic consistency, model training is completed through forward propagation and back propagation driven by a loss function, and the spatial expression and semantic consistency is improved. And performing acne target prediction on the test set by using the trained model, and outputting a detection result. According to the invention, by introducing the proxy dynamic hybrid normalization module and multi-scale feature fusion, the accuracy of acne detection and the robustness of the model to a target under a complex skin condition are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to an acne detection method and system based on proxy dynamic hybrid normalization. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology in the field of medical health, computer vision-based dermatosis assisted diagnosis has gradually become a research hotspot. Among them, facial acne, as one of the high-frequency skin problems, its automatic detection has important value for intelligent skin measurement, skin health assessment, and the formulation of personalized skin care plans. However, since acne often exhibits characteristics such as small size, uneven density, blurred edges, and low contrast with skin color in facial images, the performance of traditional object detection algorithms in such tasks still faces many challenges. The current mainstream YOLO series object detection algorithms have achieved effective recognition of targets in complex scenarios by introducing multi-scale feature fusion structures such as Feature Pyramid Network, Path Aggregation Feature Pyramid Network, and multi-way detection heads. However, these methods generally rely on fixed normalization strategies, especially Batch Normalization (BN). Its dependence on global statistics makes the model prone to normalization bias when dealing with small targets or highly inconsistent sample feature distributions, weakening the representation ability of local detail information. At the same time, although Instance Normalization (IN) has certain advantages in capturing local changes, due to its lack of global perception ability, it often fails to provide robust feature expressions. In addition, in actual detection tasks, the image acquisition environment, facial area differences, and the high diversity of acne morphology further exacerbate the uncertainty and inadaptability of the normalization strategy. Therefore, there is an urgent need to construct a normalization mechanism that can achieve a dynamic balance between instance and batch normalization and has context awareness ability to improve the accuracy and robustness of acne detection. Summary of the Invention

[0003] To solve the above problems, the present invention proposes an acne detection method and system based on proxy dynamic hybrid normalization. By using hierarchical random sampling to divide the dataset, introducing data augmentation strategies to improve sample diversity, and constructing an acne detection network model based on YOLOv11, combined with a proxy dynamic hybrid normalization module, it realizes the efficient extraction and normalization of facial acne image features, improves the robustness and detection accuracy of the model under different lighting, angles, and skin color conditions, and finally achieves accurate recognition and positioning of acne targets through multi-scale feature fusion and training optimization of the Neck module.

[0004] The specific solutions are as follows:

[0005] On the one hand, an acne detection method based on proxy dynamic hybrid normalization includes:

[0006] S1. Divide the face acne dataset into a training set and a test set by stratified random sampling;

[0007] S2. Perform data augmentation processing on the face acne data in both the training set and the test set to obtain the augmented training set and test set;

[0008] S3. Build an acne detection network model based on the YOLOv11 model. The acne detection network model includes a backbone network, a proxy dynamic hybrid normalization module, and a Neck module;

[0009] S4. Input the face acne data in the augmented training set into the acne detection network model. The backbone network extracts features from the face acne data in the training set to obtain the face acne data features in the training set. The proxy dynamic hybrid normalization module normalizes the face acne data features to obtain the normalized face acne data features. The Neck module fuses the normalized face acne data features to obtain the fused face acne data features. Based on the loss function, perform forward propagation and backward propagation training on the fused face acne data features to obtain the trained acne detection network model;

[0010] S5. Input the face acne data in the augmented test set into the trained acne detection network model for target prediction to obtain the acne detection result.

[0011] Further, the S1 specifically includes:

[0012] S11. Obtain the original face acne image data, perform multi-dimensional stratification based on the feature attributes of the original face acne image data to obtain multiple stratified subsets, and the sample feature attributes within the same layer are consistent; the feature attributes include the shooting angle, lighting condition, facial area, spatial distribution density of acne, image clarity, and skin color information of the original face acne image data; the facial area includes the forehead, nose wings, and lower jaw;

[0013] S12. Divide the stratified subset samples into a training set and a test set by a random sampling strategy with a set ratio, and control the balance of the sample quantity through a random number seed.

[0014] Further, in S4, the backbone network includes an initial convolution layer, a residual structure, and a first cross-stage partial connection unit;

[0015] The facial acne data images in the data-enhanced training set are feature extracted through the initial convolution layer, and the extracted facial acne data features are data-enhanced through the residual structure to obtain enhanced facial acne data features. The enhanced facial acne data features are feature encoded through the first cross-stage partial connection unit to extract facial acne data features containing different semantic levels and spatial resolutions for characterizing the local differences and semantic information of acne in edges, textures, and colors, and finally feature maps of different scales are obtained. .

[0016] Further, in S4, the proxy dynamic hybrid normalization module includes a dual-branch weight network and a normalization calculation unit;

[0017] The dual-branch weight network is used to receive feature maps of different scales from the backbone network. , compute feature maps of different scales through the proxy attention mechanism Based on the calculated feature correlation, the first dynamic weight is obtained through the instance normalization function, and the second dynamic weight is obtained through the batch normalization function. ; The calculation formulas for the first dynamic weight and the second dynamic weight are as follows:

[0018] ;

[0019] ;

[0020] in, represents the first dynamic weight; represents the second dynamic weight; IN represents the instance normalization function; BN represents the batch normalization function; represents the agent attention mechanism; represents the query matrix; represents the proxy matrix; represents the bond matrix; represents the value matrix; The calculation formula is as follows:

[0021] ;

[0022] ;

[0023] in, is the first learnable bias matrix; is the second learnable bias matrix; * represents the convolution operation; Represents a depthwise convolution operation; Represents a pooling operation; Represents the learnable parameter matrix corresponding to the query matrix; A parameter matrix representing the learnable matrix corresponding to the key matrix; A learnable parameter matrix representing the value matrix.

[0024] The normalization calculation unit is used to normalize the first dynamic weight and the second dynamic weight to obtain the normalized facial acne data features.

[0025] Further, the normalization calculation unit is used to normalize the first dynamic weight and the second dynamic weight to obtain the normalized facial acne data features, specifically including:

[0026] Performing convolution operations on feature maps of different scales to obtain deep features , and then linearly fusing the deep features with the first state weight and the second dynamic weight to obtain the normalized facial acne data features , and the calculation formula is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] where represents the convolution operation; represents the number of samples in each input batch; represents the height of the feature map; represents the width of the feature map; , and respectively represent the height index position, width index position and channel number index position of the feature map; and represent the mean and variance of batch normalization; and represent the mean and variance of instance normalization; represents the first affine parameter; represents the second affine parameter; Represents a constant.

[0034] Furthermore, in S4, the Neck module includes a second cross-stage partial connection unit, an upsampling unit, a downsampling unit, and a path aggregation unit;

[0035] The second cross-stage partial connection unit performs preliminary fusion processing on the normalized facial acne data features to obtain high-level features with enhanced semantic expressions; the upsampling unit aligns the high-level features with enhanced semantic expressions in the spatial dimension, and improves the feature resolution through nearest neighbor interpolation and transposed convolution to obtain aligned middle-level features; the downsampling unit performs spatial compression and channel enhancement on the aligned middle-level features, and obtains deep features with enhanced global information through convolution and pooling operations; the path aggregation unit aggregates the middle-level features and deep features to aggregate deep and shallow features, obtains the fused facial acne data features, and performs forward propagation and backpropagation training on the fused facial acne data features based on the loss function to update the network weights and obtain a trained acne detection network model.

[0036] On the other hand, an acne detection system based on proxy dynamic hybrid normalization includes:

[0037] A stratified sampling module for dividing the facial acne data set into a training set and a test set by stratified random sampling;

[0038] A data augmentation module for performing data augmentation processing on the facial acne data in both the training set and the test set to obtain an augmented training set and an augmented test set;

[0039] A model construction module for constructing an acne detection network model based on the YOLOv11 model, where the acne detection network model includes a backbone network, a proxy dynamic hybrid normalization module, and a Neck module;

[0040] A training module for inputting the facial acne data in the augmented training set into the acne detection network model, where the backbone network extracts features from the facial acne data in the training set to obtain the facial acne data features in the training set, the proxy dynamic hybrid normalization module normalizes the facial acne data features to obtain normalized facial acne data features, the Neck module fuses the normalized facial acne data features to obtain the fused facial acne data features, and performs forward propagation and backpropagation training on the fused facial acne data features based on the loss function to obtain a trained acne detection network model;

[0041] An acne detection module for inputting the facial acne data in the augmented test set into the trained acne detection network model for target prediction to obtain an acne detection result.

[0042] The present invention adopts the above technical solutions and has the following beneficial effects:

[0043] (1) By using a module based on proxy dynamic hybrid normalization to optimize the facial acne data features, and combining the dynamic weighted fusion of instance normalization and batch normalization, the present invention enhances the model's recognition ability for acne in different scales, lighting conditions, and facial regions, thereby improving the accuracy and reliability of acne detection.

[0044] (2) By using a hierarchical random sampling strategy to divide the dataset into a training set and a test set, and performing data augmentation, the present invention enables the model to learn a wider range of feature representations under diverse sample conditions. Meanwhile, the residual structure and cross-stage partial connection units in the backbone network help to maintain the continuity and integrity of the information flow, further enhancing the model's performance in the face of complex skin conditions.

[0045] (3) By integrating various mechanisms such as upsampling, downsampling, and path aggregation through the Neck module, the present invention effectively aggregates deep and shallow features, not only strengthening the spatial expression ability and semantic consistency of the features, but also being able to adapt to the detection requirements of acne targets with different sizes and distribution densities, improving the detection efficiency and performance of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flowchart of the acne detection method based on proxy dynamic hybrid normalization according to an embodiment of the present invention;

[0047] Figure 2 is an overall network structure diagram of the acne detection method based on proxy dynamic hybrid normalization according to an embodiment of the present invention;

[0048] Figure 3 is a schematic structural diagram of the proxy dynamic hybrid normalization module according to an embodiment of the present invention;

[0049] FIG. 4(a) is a diagram of the acne detection result based on proxy dynamic hybrid normalization according to an embodiment of the present invention;

[0050] FIG. 4(b) is a schematic diagram of the acne detection result in the baseline model YOLOv11 according to an embodiment of the present invention;

[0051] FIG. 4(c) is a schematic diagram of the acne detection result in the baseline model YOLOv10 according to an embodiment of the present invention;

[0052] FIG. 4(d) is a schematic diagram of the acne detection result in the baseline model YOLOv8 according to an embodiment of the present invention;

[0053] Figure 5 is a system diagram of the acne detection based on proxy dynamic hybrid normalization according to an embodiment of the present invention. Detailed implementation mode

[0054] The present invention will be further described in detail below in conjunction with embodiments and the accompanying drawings, but the implementation modes of the present invention are not limited thereto.

[0055] As Figure 1 shown, the acne detection method based on proxy dynamic hybrid normalization of the present invention includes:

[0056] S1. Divide the face acne dataset into a training set and a test set by stratified random sampling.

[0057] Specifically, the S1 specifically includes:

[0058] S11. Obtain the original face acne image data, perform multi-dimensional stratification based on the feature attributes of the original face acne image data to obtain multiple stratified subsets, and the sample feature attributes within the same layer are consistent; the feature attributes include the shooting angle, lighting conditions, facial area, spatial distribution density of acne, image clarity, and skin color information of the original face acne image data; the facial area includes the forehead, nose wings, and mandible;

[0059] S12. Divide the stratified subset samples into a training set and a test set by a random sampling strategy with a set ratio, and control the balance of the sample quantity through a random number seed.

[0060] S2. Perform data augmentation processing on the face acne data in both the training set and the test set to obtain the training set and the test set after data augmentation.

[0061] S3. Build an acne detection network model based on proxy dynamic hybrid normalization based on the YOLOv11 model. The acne detection network model includes a backbone network, a proxy dynamic hybrid normalization module, and a Neck module.

[0062] S4. Input the face acne data in the training set after data augmentation into the acne detection network model. The backbone network extracts features from the face acne data in the training set to obtain the face acne data features in the training set. The proxy dynamic hybrid normalization module normalizes the face acne data features to obtain the normalized face acne data features. The Neck module fuses the normalized face acne data features to obtain the fused face acne data features. Based on the loss function, forward propagation and backward propagation training are performed on the fused face acne data features to obtain the trained acne detection network model.

[0063] Specifically, the backbone network includes an initial convolutional layer, a residual structure, and a first cross-stage partial connection unit;

[0064] Extract features from the facial acne data images in the training set after data augmentation through the initial convolutional layer, perform data augmentation on the extracted facial acne data features through the residual structure to obtain enhanced facial acne data features, and perform feature encoding on the enhanced facial acne data features through the first cross-stage partial connection unit to extract facial acne data features containing different semantic levels and spatial resolutions for characterizing the local differences and semantic information of acne in terms of edges, textures, and colors, and finally obtain feature maps of different scales 。

[0065] Specifically, the proxy dynamic hybrid normalization module includes a two-branch weight network and a normalization calculation unit;

[0066] The two-branch weight network is used to receive feature maps of different scales from the backbone network , calculate the feature correlation between the feature maps of different scales through the proxy attention mechanism , obtain the first dynamic weight through the instance normalization function and the second dynamic weight through the batch normalization function based on the calculated feature correlation ; The calculation formulas for the first dynamic weight and the second dynamic weight are as follows:

[0067] ;

[0068] ;

[0069] where, represents the first dynamic weight; represents the second dynamic weight; IN represents the instance normalization function; BN represents the batch normalization function; represents the proxy attention mechanism; represents the query matrix; represents the proxy matrix; represents the key matrix; represents the value matrix; The calculation formula of

[0070] ;

[0071] ;

[0072] where, is the first learnable bias matrix; is the second learnable bias matrix; * represents the convolution operation; represents the depth convolution operation; represents the pooling operation; represents the learnable parameter matrix corresponding to the query matrix; A parameter matrix representing a learnable matrix corresponding to a key matrix; A learnable parameter matrix representing a value matrix.

[0073] The normalization calculation unit is used to normalize the first dynamic weight and the second dynamic weight to obtain normalized facial acne data features.

[0074] Specifically, the normalization calculation unit is used to normalize the first dynamic weight and the second dynamic weight to obtain normalized facial acne data features, specifically including:

[0075] Perform convolution operations on feature maps of different scales to obtain deep features , and then linearly fuse the deep features with the first state weight and the second dynamic weight to obtain normalized facial acne data features , and the calculation formula is as follows:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] where represents the convolution operation; represents the number of samples in each input batch; represents the height of the feature map; represents the width of the feature map; , and represent the height index position, width index position, and channel number index position of the feature map; and represent the mean and variance of batch normalization; and represent the mean and variance of instance normalization; represents the first affine parameter; represents the second affine parameter; Represents a constant.

[0083] Specifically, the Neck module includes a second cross-stage partial connection unit, an upsampling unit, a downsampling unit, and a path aggregation unit;

[0084] The second cross-stage partial connection unit performs preliminary fusion processing on the normalized facial acne data features to obtain high-level features with enhanced semantic expressions; the upsampling unit aligns the high-level features with enhanced semantic expressions in the spatial dimension, and enhances the feature resolution through nearest neighbor interpolation and transposed convolution to obtain aligned middle-level features; the downsampling unit performs spatial compression and channel enhancement on the aligned middle-level features, and obtains deep-level features with enhanced global information through convolution and pooling operations; the path aggregation unit aggregates the middle-level features and deep-level features to aggregate deep and shallow features, obtains the fused facial acne data features, and performs forward propagation and backpropagation training on the fused facial acne data features based on the loss function to update the network weights and obtain the trained acne detection network model.

[0085] S5, Input the facial acne data in the data-augmented test set into the trained acne detection network model for target prediction to obtain the acne detection result.

[0086] Specifically, in this embodiment, the publicly available facial acne dataset Acne04 is used as the input. It contains 1419 images, which are divided into a training set and a test set in a ratio of 8:2, and the divided training set and test set are used for training and testing. Data augmentation processing is performed on the obtained training set and test set, including vertical flipping, random cropping, and saturation and exposure adjustment. The image resolution adopted by the present invention is 640×640. Finally, the preprocessed training set and test set images X are obtained; then, based on the YOLOv11 model, an Agent Dynamic Hybrid Normalization (ADHN) module is designed, as Figure 2 Shown is the overall network structure diagram of the acne detection method based on agent dynamic hybrid normalization constructed in this embodiment. The preprocessed image X is sequentially passed through the Backbone network, the ADHN module, the Neck network, and the Head module. Specifically, the backbone network (Backbone) in YOLOv11 is used to extract features from the data-augmented image X to obtain three feature maps with different scales, where respectively represent the features extracted from the 5th layer (C3K2), the 7th layer (C3K2), and the 10th layer (SPPF); then as Figure 3 Shown, construct an Agent Hybrid Normalization (ADHN) module, including a two-branch weight network and a normalization calculation unit; the two-branch weight network is specifically: First, the input feature Perform a two-branch operation, respectively through a learnable query parameter matrix ( ), key parameter matrix ( ), and the value parameter matrix ( ), respectively obtain their query matrix (Q), key matrix (K), and value matrix (V). Subsequently, a pooling operation is performed on the query matrix (Q) to obtain a small group of proxy tokens (A). Subsequently, by using the proxy tokens (A) as the query, and the key matrix (K) and value matrix (V) as the key-value pairs, a Softmax attention operation is performed to obtain the proxy matrix ( ). Secondly, use the query matrix (Q) as the query, the proxy tokens (A) as the key, as the value, perform the second Softmax attention operation, and add a position bias matrix thereto. Finally, after the obtained output, a depth convolution operation is performed on the value matrix (V), and the features obtained by the double-branch network are respectively passed through the instance normalization function f(IN) and the batch normalization function f(BN) to obtain the dynamic weights and . The normalization calculation unit specifically is: perform convolution processing on the input feature to obtain the deep feature , and then linearly fuse the deep feature with the weights and to obtain the normalized acne data feature of the face . Then, through a concatenation (Concat) operation, the normalized feature is fused into the Neck network in YOLOv11, and the features extracted from the cross-stage module (C3K2) with a kernel size of 2 in the Neck network part are respectively input into the 3 detection heads (Head) in YOLOv11, and finally the corresponding acne detection results are output.

[0087] In this embodiment, the performance of the constructed model is verified. The entire model training process is iterated 200 times, and the initial learning rate is set to 0.02. After training, the visualization results of the acne detection experiment are obtained.

[0088] Figure 4(a) is a schematic diagram of the acne detection result in the method of the present invention; Figure 4(b) is a schematic diagram of the acne detection result in the baseline model YOLOv11; Figure 4(c) is a schematic diagram of the acne detection result in the comparison method YOLOv10; Figure 4(d) is a schematic diagram of the acne detection result in the comparison method YOLOv8. From Figures 4(a) to 4(d) the schematic diagram of the detection result, it can be seen that the method of the present invention can detect and identify more acne targets compared to other models.

[0089] Generally speaking, the acne detection model constructed by this method can improve the recognition ability of small targets in face acne detection. On the basis of the original YOLOv11 object detection model, an Agent Dynamic Hybrid Normalization module (ADHN) is introduced to solve the fusion and adaptability problems between instance normalization and batch normalization. First, the input image passes through the Backbone network to extract multi-scale feature maps. Subsequently, through a two-branch weight network constructed by the proxy attention mechanism, the statistical features (mean and variance) of instance normalization and batch normalization are extracted respectively, and linearly fused through adaptive weight dynamics to generate a more discriminative comprehensive normalization feature map. This feature fusion process further aggregates multi-scale information in the Neck part by combining the PAFPN structure, thereby enhancing the model's response to features in different distribution regions. Finally, the fused feature map is input into the Head detection head for object classification and regression prediction to obtain the final detection result. The experimental verification results on the public dataset Acne04 show that the proposed method has better detection accuracy in acne targets, especially in low-density and small-target regions, effectively solving problems such as false detection and missed detection of traditional methods. This method is not only applicable to the acne detection task, but also provides a general normalization fusion design idea for other small-target detections, with good promotion value.

[0090] As Figure 5 shown, this embodiment also discloses an acne detection system based on Agent Dynamic Hybrid Normalization, including:

[0091] A stratified sampling module 51, used to divide the face acne dataset into a training set and a test set through stratified random sampling;

[0092] A data augmentation module 52, used to perform data augmentation processing on the face acne data in both the training set and the test set to obtain an augmented training set and an augmented test set;

[0093] A model construction module 53, used to construct an acne detection network model based on the YOLOv11 model, where the acne detection network model includes a backbone network, an Agent Dynamic Hybrid Normalization module, and a Neck module;

[0094] A training module 54, used to input the face acne data in the augmented training set into the acne detection network model. The backbone network extracts features from the face acne data in the training set to obtain the face acne data features in the training set. The Agent Dynamic Hybrid Normalization module performs normalization processing on the face acne data features to obtain normalized face acne data features. The Neck module fuses the normalized face acne data features to obtain the fused face acne data features, and performs forward propagation and backward propagation training on the fused face acne data features based on the loss function to obtain a trained acne detection network model;

[0095] The acne detection module 55 is configured to input the facial acne data in the data-augmented test set into the trained acne detection network model for target prediction to obtain acne detection results.

[0096] The specific implementation of the acne detection system based on proxy dynamic hybrid normalization is the same as that of the acne detection method based on proxy dynamic hybrid normalization, and will not be repeated in this embodiment.

[0097] Although the present invention has been specifically shown and described in conjunction with the preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made to the present invention without departing from the spirit and scope of the present invention defined by the appended claims, and all of them fall within the protection scope of the present invention.

Claims

1. A method for acne detection based on proxy dynamic hybrid normalization, characterized in that, Including: S1. Divide the facial acne dataset into a training set and a test set by stratified random sampling; S2. Perform data augmentation on the facial acne data in both the training set and the test set to obtain the augmented training set and test set; S3. Build an acne detection network model based on the YOLOv11 model. The acne detection network model includes a backbone network, a proxy dynamic hybrid normalization module, and a Neck module; S4. Input the facial acne data in the augmented training set into the acne detection network model. The backbone network extracts features from the facial acne data in the training set to obtain the facial acne data features in the training set. The proxy dynamic hybrid normalization module normalizes the facial acne data features to obtain the normalized facial acne data features. The Neck module fuses the normalized facial acne data features to obtain the fused facial acne data features. Perform forward propagation and backpropagation training on the fused facial acne data features based on the loss function to obtain the trained acne detection network model; S5. Input the facial acne data in the augmented test set into the trained acne detection network model for target prediction to obtain the acne detection result.

2. The acne detection method based on proxy dynamic hybrid normalization according to claim 1, characterized in that, The S1 specifically includes: S11. Obtain the original facial acne image data, perform multi-dimensional stratification based on the characteristic attributes of the original facial acne image data to obtain multiple stratified subsets, and the sample characteristic attributes within the same level are consistent; the characteristic attributes include the shooting angle, lighting condition, facial area, spatial distribution density of acne, image clarity, and skin color information of the original facial acne image data; the facial area includes the forehead, nose wing, and mandible; S12. Divide the stratified subset samples into a training set and a test set by a random sampling strategy with a set ratio, and control the balance of the sample quantity through a random number seed.

3. The acne detection method based on proxy dynamic hybrid normalization according to claim 1, characterized in that In S4, the backbone network includes an initial convolution layer, a residual structure, and a first cross-stage partial connection unit; Feature extraction is performed on the facial acne data images in the training set after data augmentation through the initial convolutional layer. Data augmentation is performed on the extracted facial acne data features through the residual structure to obtain enhanced facial acne data features. Feature encoding is performed on the enhanced facial acne data features through the first cross-stage partial connection unit to extract facial acne data features containing different semantic levels and spatial resolutions for characterizing the local differences and semantic information of acne in terms of edges, textures, and colors, and finally feature maps of different scales are obtained. 。 4. The acne detection method based on proxy dynamic hybrid normalization according to claim 3, characterized in that, In S4, the proxy dynamic hybrid normalization module includes a double-branch weight network and a normalization calculation unit; The dual-branch weight network is used to receive feature maps of different scales from the backbone network , and calculate the feature correlations between the feature maps of different scales through the proxy attention mechanism . Based on the calculated feature correlations, the first dynamic weight is obtained through the instance normalization function, and the second dynamic weight is obtained through the batch normalization function ; The calculation formulas for the first dynamic weight and the second dynamic weight are as follows: ； ； Among them, represents the first dynamic weight; represents the second dynamic weight; IN represents the instance normalization function; BN represents the batch normalization function; represents the proxy attention mechanism; represents the query matrix; represents the proxy matrix; represents the key matrix; represents the value matrix; The calculation formula of is as follows: ； ； Among them, is the first learnable bias matrix; is the second learnable bias matrix; * represents the convolution operation; represents the depth convolution operation; represents the pooling operation; represents the learnable parameter matrix corresponding to the query matrix; represents the parameter matrix of the learnable matrix corresponding to the key matrix; represents the learnable parameter matrix corresponding to the value matrix; The normalization calculation unit is used to normalize the first dynamic weight and the second dynamic weight to obtain the normalized facial acne data features.

5. The acne detection method based on agent dynamic hybrid normalization according to claim 4, wherein The normalization calculation unit is used to perform normalization processing on the first dynamic weight and the second dynamic weight to obtain normalized facial acne data features, specifically including: Feature maps of different scales are subjected to convolution operations to obtain deep features . Then, the deep features are linearly fused with the first-state weight and the second dynamic weight to obtain normalized facial acne data features . The calculation formula is as follows: ； ； ； ； ； ； Among them, represents a convolution operation; represents the number of samples per input batch; represents the height of the feature map; represents the width of the feature map; , and respectively represent the height index position, width index position, and channel number index position of the feature map; and represent the mean and variance of batch normalization; and represent the mean and variance of instance normalization; represents the first affine parameter; represents the second affine parameter; represents a constant.

6. The acne detection method based on proxy dynamic hybrid normalization according to claim 1, characterized in that In S4, the Neck module includes a second cross-stage partial connection unit, an upsampling unit, a downsampling unit, and a path aggregation unit; The second cross-stage partial connection unit performs preliminary fusion processing on the normalized facial acne data features to obtain high-level features with enhanced semantic expression; The upsampling unit aligns the high-level features with enhanced semantic expression in the spatial dimension, and improves the feature resolution through nearest neighbor interpolation and transposed convolution to obtain the aligned middle-level features; The downsampling unit performs spatial compression and channel enhancement on the aligned middle-level features, and obtains deep features with enhanced global information through convolution and pooling operations; The path aggregation unit aggregates the middle-level features and the deep features to aggregate the deep and shallow features to obtain the fused facial acne data features. Perform forward propagation and backpropagation training on the fused facial acne data features based on the loss function to update the network weights and obtain the trained acne detection network model.

7. An acne detection system based on proxy dynamic hybrid normalization, characterized in that, Including: A stratified sampling module for dividing the facial acne dataset into a training set and a test set through stratified random sampling; A data augmentation module for performing data augmentation processing on the facial acne data in both the training set and the test set to obtain the augmented training set and test set; A model construction module for constructing an acne detection network model based on the YOLOv11 model, where the acne detection network model includes a backbone network, a proxy dynamic hybrid normalization module, and a Neck module; A training module for inputting the facial acne data in the augmented training set into the acne detection network model. The backbone network extracts features from the facial acne data in the training set to obtain the facial acne data features in the training set. The proxy dynamic hybrid normalization module normalizes the facial acne data features to obtain the normalized facial acne data features. The Neck module fuses the normalized facial acne data features to obtain the fused facial acne data features. Based on the loss function, forward propagation and backward propagation training are performed on the fused facial acne data features to obtain the trained acne detection network model; An acne detection module for inputting the facial acne data in the augmented test set into the trained acne detection network model for target prediction to obtain the acne detection result.

Images