Cross-channel semantic segmentation method for generating pseudo mode based on wavelet transform
By generating pseudo-modal data through wavelet transform and combining it with an improved cross-attention UNet model, the problems of insufficient data mining and semantic gap between encoder and decoder in semantic segmentation are solved, achieving higher precision and more accurate semantic segmentation.
Patent Information
- Application Number
- CN202411009379.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-01
AI Technical Summary
Existing semantic segmentation methods have shortcomings in data mining and feature extraction, resulting in limited segmentation accuracy and generalization ability. In particular, the semantic gap between the encoder and decoder has not been effectively addressed.
Pseudomodal data is generated using wavelet transform and combined with an improved cross-attention UNet model. Feature cross-fusion is enhanced by projection and activation feature interaction attention module and dual fusion cross-attention module. The Dice Loss loss function is used for training to improve segmentation performance.
By utilizing diverse data mining and information complementarity, the accuracy of semantic segmentation and edge localization was improved, the semantic gap between the encoder and decoder was resolved, and segmentation performance was enhanced.
Smart Images

Figure CN118887403A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image processing, computer vision and semantic segmentation models, and in particular relates to a cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality. Background Art
[0002] Semantic segmentation is a task in computer vision, which aims to divide an image into semantically meaningful regions and classify each pixel. Specifically, semantic segmentation classifies each pixel in an image into a specific object category, such as a person, a car, a tree, etc. In this way, not only can different objects in the image be identified, but the position and boundaries of these objects can also be accurately located. Unlike object detection, semantic segmentation can provide more fine-grained image understanding, identify the category to which each pixel belongs, and thus obtain more detailed information about the image content. Semantic segmentation has important application value in many fields of social production. In autonomous driving, semantic segmentation is used to identify roads, pedestrians, vehicles, and other traffic elements, helping autonomous driving systems to better understand and analyze the surrounding environment and make safe decisions. In medical image analysis, semantic segmentation can be used to segment and mark different organs and lesion areas, which helps doctors make diagnoses and treatment plans. Semantic segmentation can be used for image editing, such as background replacement and object removal, and precise pixel-level segmentation makes editing more natural and realistic. In smart city management, semantic segmentation of camera images can be used to achieve applications such as traffic flow analysis, public facility management, and security monitoring. In short, semantic segmentation is of great significance in improving computers' ability to understand images and promoting practical applications.
[0003] Early semantic segmentation methods were mainly based on traditional image processing techniques, such as edge detection and region growing. These methods mainly rely on low-level features and hand-designed rules to achieve image segmentation, with limited accuracy and generalization ability. With the rise of deep learning, semantic segmentation has entered a new stage based on neural networks. The first is the method based on fully convolutional networks, which applies convolutional neural networks (CNNs) to pixel-level classification tasks to achieve more accurate semantic segmentation. Subsequently, a series of improved network architectures emerged, such as U-Net, SegNet, DeepLab, etc. These networks have been optimized in structural design and loss functions to improve segmentation accuracy and speed. Although the emergence of deep learning has improved the accuracy of semantic segmentation, most of them remain at the improvement of the model, and there are still few studies on mining the data itself to improve segmentation performance. Summary of the invention
[0004] In view of the above problems, the present invention proposes a cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality to improve segmentation performance.
[0005] A cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality includes the following steps:
[0006] S1. Obtain segmented image data and corresponding labels.
[0007] S2. Preprocess the obtained image data. The preprocessing steps are as follows:
[0008] S21, normalize the image data, reduce the value of the image data to between 0 and 1, and adjust the size of the image and label data to a uniform size.
[0009] S22. For experimental images and their annotated images, image data enhancement algorithm is used to expand the image dataset.
[0010] S3, performing wavelet transform on the pre-processed image data to obtain a child image as pseudo-modal data. The specific steps are as follows:
[0011] S31. First, use Haar basis to perform wavelet transform to obtain 4 subband data.
[0012] S32, resizing the sub-band data generated by wavelet transform.
[0013] S4, sorting the preprocessed image data and the generated pseudo-modality data to form a new data set. The specific steps are as follows:
[0014] S41. Arrange the preprocessed image data and the generated pseudo-modal data by channels to form new multi-channel data, constituting a new data set.
[0015] S42. The newly generated data set is randomly divided into a training set and a test set in proportion, wherein a 5-fold cross validation is performed on the training set.
[0016] S5. Improve the cross-attention UNet model to build a cross-channel semantic segmentation model, input the training set data into the cross-channel semantic segmentation model for training, output the semantic segmentation results, and then input the test set data into the trained converged model for segmentation to verify the model effect. The specific steps are as follows:
[0017] S51. In order to obtain the cross-complementary information of the original data and the pseudo-modal data, the encoding module of the encoding part is replaced by the projection and activation feature interactive attention module proposed in the present invention on the original encoding and decoding structure of UNet, so as to obtain cross information and improve the segmentation accuracy. At the same time, the dual fusion cross attention module proposed in the present invention is added to the jump connection part of the model.
[0018] S511. The projection and activation feature interaction attention module proposed in the present invention is mainly composed of an input stage, a feature interaction and fusion stage, and an output stage. The input stage mainly divides the input data set into two feature tensors by channel. The feature interaction and fusion stage is mainly composed of two feature interaction branches and a projection activation attention module, wherein the two feature interaction branches are used to extract features from the two feature tensors divided in the input stage, and to adjust the feature weights of the interaction weights obtained by the activation mapping cross attention module. The activation mapping cross attention module is used to perform multi-dimensional feature mapping and activation on the complete input tensor to obtain the feature weight coefficients corresponding to the two feature branches. Finally, the output stage will perform channel cascading on the output and input tensors of the two feature branches and adjust the shape of the tensors.
[0019] S512. The core of the dual fusion cross attention module proposed in the present invention is composed of a channel cross attention module and a spatial cross attention module. The two modules process the input data in a parallel residual connection manner, and finally add, normalize and activate the features extracted from the two parts to obtain the final features.
[0020] S52. Use the Dice Loss function to train the cross-channel semantic segmentation model, making the training effect of the cross-channel semantic segmentation model more effective.
[0021] S53. Input the original data and pseudo-modal data in the test set into the trained converged cross-channel semantic segmentation model for semantic segmentation, and verify the segmentation results.
[0022] Compared with the prior art, the present invention has the following beneficial effects:
[0023] (1) This method uses wavelet transform to obtain pseudo-modal data and further mines the characteristics of the original data itself.
[0024] Subsequently, the pseudo-modal data and the original data are fused and trained in the model. Compared with the traditional training using only the original data, the model can obtain more favorable features during training. Through the cross-information of the two types of data, the model can locate the segmented objects more accurately and the segmentation edges will be more precise.
[0025] (2) The improved cross-attention UNet model extracted by this method, on the one hand, fits the diversity of input data. The added projection and activation feature interaction attention modules cross-fuse the original data and pseudo-modal data multiple times in the encoding stage to extract more semantic information. Compared with the existing semantic segmentation model, in the process of feature extraction of each encoding module, only the overall situation of the input data is paid attention to, but the internal information interaction is not paid attention to. Therefore, the model proposed by this method will be more flexible and efficient in feature extraction. On the other hand, most of the current UNet improved models will add skip connections between the encoder and decoder to improve the problem of spatial information loss and semantic information gap. However, this method is too simple to deal with the problem. After a simple skip connection, the model segmentation performance is still not accurate enough because of the semantic gap between the encoder and decoder. The dual fusion cross attention module proposed in this method exchanges information between the encoder and decoder in terms of space and channels, solves the problem of semantic gap, makes the information of the skip connection more equal, and improves the segmentation performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A flowchart of the cross-attention universal semantic segmentation method for generating pseudo-modality based on wavelet transform of the present invention;
[0027] Figure 2 The process of performing wavelet transformation on the original data in step S3 of the present invention;
[0028] Figure 3 The network structure diagram of the improved cross-attention UNet algorithm proposed in the present invention;
[0029] Figure 4 The network structure diagram of the projection and activation feature interactive attention module proposed in the present invention;
[0030] Figure 5 This is the network structure diagram of the dual fusion cross attention module proposed in the present invention;
[0031] Figure 6 The result diagram of the training set and the validation set of the improved model proposed in the present invention;
[0032] Figure 7 This is a diagram of the prediction effect of the improved network model proposed in the present invention. DETAILED DESCRIPTION
[0033] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0034] It should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs. The terms used in the present invention and in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.
[0035] like Figure 1 As shown, taking three-dimensional medical data as an example, a cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality includes the following steps:
[0036] S1. Arrange the data and organize the segmented image data and corresponding annotation labels.
[0037] S2. Preprocess the obtained image data. The preprocessing steps are as follows:
[0038] S21. If the data is two-dimensional image data, first, convert the image data and label data into PNG format. Secondly, normalize the original image data, reduce the value of the image data to between 0 and 1, set the background part of the label data to 0, and set the segmented object part to 1. Finally, adjust the size of the image and label data to 512×512.
[0039] S22. If the data is 3D image data, first, convert the image data and label data into nii.gz format. Secondly, normalize the original image data, reduce the value of the image data to between 0 and 1, set the background part of the label data to 0, and set the segmented object part to 1. Finally, adjust the size of the image and label data to 64×64×64.
[0040] S23. For experimental images and their annotated images, a specific image data enhancement algorithm is used to expand the image dataset.
[0041] In one embodiment, the method for performing data augmentation and data set expansion on the experimental images and their annotated images includes continuing to perform gamma correction, as well as random flipping (up and down, left and right) and random rotation angles (between 45° and 90°) each time.
[0042] S3, performing wavelet transform on the pre-processed image data to obtain a child image as pseudo-modal data. The specific steps are as follows:
[0043] S31. First, use Haar basis to perform wavelet transform to obtain four sub-band data, namely Approximation (LL), Horizontal detail (LH), Vertical detail (HL), Diagonal detail (HH). The size of the four sub-band data is half of the original data. Figure 2 shown.
[0044] S32, resizing the sub-band data generated by wavelet transform, selecting Vertical detail (HL) as pseudo-modal data, and resizing it to the size of the original data.
[0045] In one example, Approximation (LL) represents the approximation coefficient of the low-frequency part, which represents the overall structure and general outline of the image. Horizontal detail (LH) represents the high-frequency detail coefficient in the horizontal direction, which captures the horizontal changes and fine structures in the image. Vertical detail (HL) represents the high-frequency detail coefficient in the vertical direction, which reflects the vertical changes and details in the image. Diagonal detail (HH) represents the high-frequency detail coefficient in the diagonal direction, which is used to capture the diagonal changes and fine structures in the image.
[0046] S4, sorting the preprocessed image data and the generated pseudo-modality data to form a new data set. The specific steps are as follows:
[0047] S41. Arrange the original data and the generated pseudo-modal data by channels to form new multi-channel data, which constitute a new data set.
[0048] S42. The newly generated data set is randomly divided into a training set and a test set in a ratio of 8:2. The training set is subjected to 5-fold cross-validation, and the test set is used to verify the model effect.
[0049] S5. Improve the cross-attention UNet model to build a cross-channel semantic segmentation model, input the training set data into the cross-channel semantic segmentation model for training, and output the semantic segmentation results. Subsequently, input the data in the test set into the trained converged model for segmentation to verify the model effect. The specific steps are as follows:
[0050] S51. In order to obtain the cross-complementary information of the original data and the pseudo-modal data, the encoding module of the encoding part is replaced with the projection and activation feature interactive attention module proposed in the present invention to obtain the cross information. The double fusion cross attention module proposed in the present invention is added to the jump connection part to solve the semantic gap between the encoder and the decoder and improve the segmentation accuracy. The network structure diagram of the improved cross attention UNet algorithm is shown in the figure below: Figure 3 As shown in the figure, the overall structure is a U-shaped structure, which consists of an encoder, a bottleneck layer, a decoder and a jump connection. The encoder consists of 4 encoding modules, and the first three encoding modules are composed of projection and activation feature interactive attention modules, convolution modules and maximum pooling. The bottleneck layer is composed of two convolution modules. The decoder is also composed of 4 decoder modules, and the 4 decoder modules are composed of two convolution modules and an upsampling module. The jump connection part adds the dual fusion cross attention module proposed by this method.
[0051] S511. The projection and activation feature interactive attention module proposed in the present invention is as follows Figure 4 As shown in the figure, the module is divided into three stages: input, interaction and fusion of features, and output.
[0052] In the input stage: the dimension size of the feature input is Where C is the channel, D is the length, W is the width, and H is the height. The input feature tensor is divided into Two groups of tensors are symmetrical by channel. On the one hand, dividing the input tensor into two branches can significantly reduce the computational cost in the subsequent feature extraction process. On the other hand, the features in the same channel are consistent, and the features in different channels are different and complementary. Grouping the input features for learning can obtain effective specific information of multiple channels and provide a basis for the subsequent feature interaction and fusion process.
[0053] Feature interaction and fusion stage: In the input stage, the input tensor is divided into two branches. In this stage, there are two parallel 3D convolution paths to extract features from the input tensor respectively, and obtain multi-channel information within the two branches. Input into feature path 1 and feature path 2, and calculated by the convolution module The calculation process is as follows:
[0054]
[0055] Among them, Relu represents the activation function, BN represents batch normalization, and 3D Conv 3×3×3 represents a 3D convolution module with a convolution kernel of 3. In order to promote information interaction between different pathways, the input of pathway 1 is cascaded across branches. and Perform the Concat cascade operation on the channel to obtain Input of Path 2 and Perform Concat cascade operation on the channel to obtain The specific expressions are as follows:
[0056]
[0057] Concat means to perform cascade operation on channels. Directly performing cascade operation on channels will undoubtedly add a lot of redundant information. Therefore, a projection activation attention module is designed to adaptively correct the result of the two-channel cascade. Then, projection operations are performed at three angles: D, W, and H. First, the maximum pooling operation is performed in three dimensions to obtain The specific expressions are as follows:
[0058]
[0059] Among them, 3DMaxPooling(·) D ,3DMaxPooling(·) W ,3DMaxPooling(·) H Respectively represent the 3D maximum pooling operation, and retain the original dimension in the subscript dimension. The middle dimension is 1, and the parts are compressed and added together to obtain The specific expressions are as follows:
[0060]
[0061] Sum(·) represents the addition operation, and squeeze(·) represents the dimension compression operation. To activate, first Transformed into a one-dimensional tensor for serialization operations In order to reduce information redundancy and huge computational overhead, Perform dimensionality reduction to obtain a tensor representing global information Subsequent Use Sigmoid for activation, and we get The specific expressions are as follows:
[0062]
[0063] Among them, Reshape(·) represents the operation of flattening multidimensional data into a one-dimensional serialized tensor, FC(·) represents the full connection operation, which transforms the length of the serialized data, and Sigmoid(·) represents the operation of using the sigmoid activation function. Perform FC1 and FC2 operations respectively, and then perform Softmax operation to obtain the global adaptive weight calibration tensor corresponding to the two feature paths. The specific expressions are as follows:
[0064]
[0065]
[0066] Among them, FC1(·), FC2(·) represent two linear functions with different weights, and Sm(·) represents the normalized soft maximum function. Finally, it represents two feature paths and After α and β are used to correct the weights, we get and The specific expressions are as follows:
[0067]
[0068] Where ⊙ represents element-wise multiplication.
[0069] Output stage: Considering and Contains two feature pathways with rich cross-information features. Contains the original global information features, so and By connecting them in series by channel, we can get a tensor rich in cross information and global information, and perform convolution, normalization and activation operations on the tensor to get the output tensor The specific operations are as follows:
[0070]
[0071] S512. The dual fusion cross attention module proposed in the present invention is shown in 5. The dual fusion cross attention module applied to 2D semantic segmentation is improved to obtain a dual fusion cross attention module that is more suitable for research work. The main improvements are as follows:
[0072] Improve the module adapted to 2D semantic segmentation network to also adapt to 3D semantic segmentation network: Change the dual cross attention module adapted to 2D network to the dual fusion cross attention module adapted to 3D network, mainly in the stage of generating Token blocks. Its patch sizes become The method of extracting patches is changed to using 3D average pooling, and the method of projecting patches to obtain tokens is changed to using a 3D convolution module with a convolution kernel of 1. The specific representation is as follows:
[0073]
[0074] Among them, 3D Conv 1×1×1 represents a 3D convolution operation with a convolution kernel of 1, Reshape represents a transformation operation on the shape of the tensor, 3DAvgPooling represents a 3D average pooling operation, and T i Represents the Token block obtained by projection.
[0075] Adjustment of the layout of channel cross-attention module and spatial cross-attention module:
[0076] In view of the need for semantic segmentation in the design of the present invention, the layout of the channel cross attention module and the spatial cross attention module in the dual cross attention module is changed, such as Figure 5 shown.
[0077] In designing the dual-fusion cross-attention module, first, the input tensor The channel criss-cross attention module CCA and the spatial criss-cross attention module CCA are executed in parallel, and Residual connections are made with the output tensors of the channel cross-attention module and the spatial cross-attention module respectively. Finally, the two residual connected tensors are added together to obtain the final The specific expressions are as follows:
[0078]
[0079] in, Represents element-wise addition.
[0080] S52. Use the Dice Loss loss function to train the cross-channel semantic segmentation model to make model training more efficient. The data-expanded training set will be used to train the improved cross-attention UNet semantic segmentation model, and the training process will use Dice Loss as the loss function. After each training epoch, save the current model weights. After each batch of training, use the validation set to evaluate the generalization performance of the model, and adjust the learning rate according to the total training loss and validation set loss values. After all the training is completed, use the test set to verify the final generalization performance of the model. Next, adjust the hyperparameters according to the characteristics of the model and data, and finally save the weight file with the smallest total loss and validation set loss as the final model weight parameters.
[0081] The specific steps are as follows: Use the PyTorch deep learning framework to build an improved cross-attention UNet semantic segmentation model. Select Dice Loss as the main loss function and use the Adam optimizer to optimize the model. The initial learning rate is set to 0.001. During the training process, the learning rate is adjusted using cosine annealing. Input the expanded image data into the model and train for 200 epochs. Set the Batch Size to 24. Save the model parameters at the end of each epoch and use the early stopping strategy. If the accuracy on the validation set has not changed in 20 consecutive trainings, stop training to prevent overfitting. Finally, save the weight file with the smallest total loss and validation set loss during training as the final weight parameter of the model. Through the above process, ensure the training and tuning of the model, and finally minimize the training set and validation set loss values of the model, such as Figure 6 As shown, the best generalization performance is achieved.
[0082] S53, input the original data and pseudo-modal data in the test set into the trained converged model for semantic segmentation, and verify the segmentation results. Figure 7 shown.
Claims
1. A cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality, It is characterized in that The steps include: S1, obtain segmented image data and corresponding labels; S2, preprocessing the obtained image data; S3, performing wavelet transform on the preprocessed image data to obtain a descendant image as pseudo-modal data; S4, forming a new data set from the preprocessed image data and the generated pseudo-modality data, and dividing it into a training set and a test set; S5. Improve the cross-attention UNet model to build a cross-channel semantic segmentation model, input the training set data into the cross-channel semantic segmentation model for training, and output the semantic segmentation results.
2. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 1, It is characterized in that The specific process of the pretreatment is as follows: S21, normalize the image data, reduce the value of the image data to between 0 and 1, and adjust the size of the image and label data to a uniform size; S22. Use image data enhancement algorithm to expand the image data set.
3. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 1, It is characterized in that The specific process of step S3 is as follows: S31, using Haar basis to perform wavelet transform to obtain 4 subband data; S32, resizing the sub-band data generated by wavelet transform.
4. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 3, It is characterized in that The specific process of forming a new data set is as follows: S41, arranging the preprocessed image data and the generated pseudo-modality data by channels to form new multi-channel data to constitute a new data set; S42. The newly generated data set is randomly divided into a training set and a test set in proportion, wherein a 5-fold cross validation is performed on the training set.
5. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 4, It is characterized in that The cross-channel semantic segmentation model is specifically implemented as follows: S51. In the original encoding and decoding structure of UNet, the encoding module of the encoding part is replaced with the projection and activation feature interactive attention module to obtain cross information, and a dual fusion cross attention module is added to the jump connection part; S52. Use the Dice Loss function to train the cross-channel semantic segmentation model.
6. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 5, It is characterized in that The projection and activation feature interaction attention module consists of an input stage, a feature interaction and fusion stage, and an output stage; The input stage: the dimension size of the feature input is Where C is the channel, D is the length, W is the width, and H is the height. The input feature tensor is divided into Two groups of channel-symmetric tensors; The interaction and fusion stage of the features: There are two parallel 3D convolution paths that extract features from the input tensor and adjust the feature weights to obtain two output features The output stage: and Concatenate by channel to get a tensor containing cross information and global information, and perform convolution, normalization and activation operations on the tensor to get the output tensor 7. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 6, It is characterized in that The specific implementation process of the interaction and fusion stage of the features is as follows: Will Input into feature path 1 and feature path 2, and calculated by the convolution module The calculation process is as follows: Among them, Relu represents the activation function, BN represents batch normalization, and 3D Conv 3×3×3 Represents a 3D convolution module with a convolution kernel of 3; The input of channel 1 is connected by cross-branch cascade. and Perform the Concat cascade operation on the channel to obtain Input of Path 2 and Perform the Concat cascade operation on the channel to obtain Design a projection activation attention module to adaptively correct the results of the two-path cascade: Then, projection operations are performed at three angles: D, W, and H. First, the maximum pooling operation is performed in three dimensions to obtain Among them, 3DMaxPooling(·) D ,3DMaxPooling(·) W ,3DMaxPooling(·) H Respectively represent the 3D maximum pooling operation and retain the original dimension in the subscript dimension; The middle dimension is 1, and the parts are compressed and added together to obtain The specific expressions are as follows: Sum(·) represents the addition operation, and squeeze(·) represents the dimension compression operation; Secondly, To activate: Transformed into a one-dimensional tensor for serialization operations right Perform dimensionality reduction to obtain a tensor representing global information right Use Sigmoid for activation, and we get Finally FC 1 and FC 2 Operation, followed by Softmax operation to obtain the global adaptive weight calibration tensor corresponding to the two feature paths Among them FC 1 (·),FC 2 (·) represents two linear functions with different weights, Sm(·) represents the normalized soft maximum function, representing two feature paths and After α and β are used to correct the weights, we get and The specific expressions are as follows: Where ⊙ represents element-wise multiplication.
8. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 5, It is characterized in that The specific implementation process of the dual fusion cross attention module is as follows: The input tensor The channel criss-cross attention module CCA and the spatial criss-cross attention module CCA are executed in parallel, and Residual connections are made with the output tensors of the channel cross-attention module and the spatial cross-attention module respectively. Finally, the two residual connected tensors are added together to obtain the final The specific expressions are as follows: in, Represents element-wise addition.
9. The cross-channel semantic segmentation method based on wavelet transform to generate pseudo-modality according to claim 1, It is characterized in that The step S5 also includes inputting the data in the test set into the cross-channel semantic segmentation model that has been trained and converged for segmentation to verify the model effect.
Citation Information
Cited By
Image crack identification model generation method for identifying road surface cracks, and application and system thereof
CN122695488A
