Stratus abdominis ultrasonic image enhancement method for postpartum rehabilitation optimization

Abstract

The invention provides a rectus abdominis ultrasonic image enhancement method for postpartum rehabilitation optimization, and relates to the field of image enhancement, a multi-dimensional feature enhancement network proposed by the invention combines a CNN and a Transform to realize efficient image enhancement, an encoder and a decoder are processed by using a CNN module, a Transform module is introduced into a global context feature extraction module, and the image enhancement efficiency is improved. The advantages of CNN in the aspect of local feature modeling and the capability of Transform in global feature extraction are fully utilized, multiple attention mechanisms are captured by designing an adaptive convolution block, and the local feature extraction effect is further enhanced, so that good balance is achieved between the image enhancement performance and the calculation complexity.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a rectus abdominis ultrasound image enhancement method for postpartum rehabilitation optimization. Background Art

[0002] As the importance of postpartum rehabilitation gradually increases, rectus abdominis ultrasound examination has become an important means to evaluate maternal rectus abdominis separation and rehabilitation progress. Postpartum rectus abdominis separation is a common condition, especially in the late pregnancy, because the abdominal wall muscles are separated by the pressure of the increased fetal volume, resulting in abdominal muscle relaxation and mechanical dysfunction. Ultrasound imaging is widely used to diagnose and evaluate this pathological change. However, ultrasound images face several challenges in their application, mainly including image quality interference by noise and artifacts, blurred edges, and insufficient resolution of tissue structure. These problems often make it difficult to accurately identify the edges of the rectus abdominis, affecting the accurate assessment of the degree of rectus abdominis separation, and may also cause inaccurate judgments in subsequent rehabilitation treatments. Therefore, how to enhance the quality of rectus abdominis ultrasound images, make the edges of the rectus abdominis clearer, and improve the contrast and resolution of tissues is an important research topic.

[0003] In recent years, the application of convolutional neural network (CNN) and Transformer-based models in the field of computer vision has gradually increased, especially in image enhancement, super-resolution, and object detection tasks. However, CNN-based models have defects in capturing long-range dependencies in images. Transformer-based models are suitable for capturing long-range dependencies in images, but compared with CNN-based models, Transformer-based models are less efficient in extracting local information. Therefore, CNN and Transformer need to be used in combination. Recently, U-Net architecture combines CNN and Transformer to seek a good balance between performance and computational complexity in various tasks. A good balance is achieved. However, where and how CNN and Transformer are effectively and efficiently integrated into the U-Net architecture requires further exploration. In view of this, the present invention proposes a rectus abdominis ultrasound image enhancement method for postpartum rehabilitation optimization. The method uses CNN to extract local features in the encoder and decoder. At the same time, in order to improve the capacity of CNN blocks without sacrificing too much efficiency, CNN-based adaptive convolution blocks are designed in the encoder and decoder to learn the dynamic complementary attention of convolution kernels. A Transformer-based global context feature extraction module is used between the encoder and decoder to extract global features, which improves the model efficiency while enhancing the rectus abdominis ultrasound image for postpartum rehabilitation optimization. Summary of the invention

[0004] The present invention provides a rectus abdominis ultrasound image enhancement method for postpartum rehabilitation optimization, aiming to achieve efficient image enhancement by constructing a multidimensional feature enhancement network, combining CNN and Transformer, using CNN modules to process encoders and decoders, and placing the Transformer module in a global context feature extraction module to give full play to the local feature extraction capabilities of CNN and the global feature extraction capabilities of Transformer, designing an adaptive convolution block to learn the multiple attentions of the convolution kernel, further improving the effect of local feature extraction, and ensuring the balance between performance and computational complexity of image enhancement.

[0005] The present invention aims to propose a rectus abdominis ultrasound image enhancement model and a rectus abdominis ultrasound image enhancement method for postpartum rehabilitation optimization, comprising the following steps.

[0006] S1. Obtain the image to be processed and use a high-frequency probe to obtain an ultrasound image of the rectus abdominis muscle for postpartum rehabilitation optimization.

[0007] S2. Construct an adaptive convolution block, including an average pooling layer, a fully connected layer, a GeLU activation function, and a traditional convolution, to learn the attention of the three dimensions of space, channel, and edge in image features and effectively aggregate local context information.

[0008] S3. Construct a dynamic local attention module, including batch normalization layer, local self-attention and adaptive convolution blocks, to extract local information.

[0009] S4. Build a global context feature extraction module, including layer normalization, dilated convolution, deep convolution layer and Softmax function, to extract global context information.

[0010] S5. Construct a multi-dimensional feature enhancement network, which consists of an encoder, a global context feature extraction module, and a decoder. The encoder consists of a dynamic local attention module and downsampling, and the decoder consists of an upsampling and dynamic local attention module.

[0011] S6. Construct a rectus abdominis ultrasound image enhancement model, which consists of a shallow feature extraction module, a multidimensional feature recovery network and an image reconstruction module. Input the rectus abdominis ultrasound image optimized for postpartum rehabilitation into the model to obtain an enhanced rectus abdominis ultrasound image optimized for postpartum rehabilitation.

[0012] Preferably, in step S2, for the adaptive convolution block, the input image feature , using an average pooling layer to reduce Dimension, reducing the parameters and computational complexity of the model. , Represents the average pooling operation. The pooling result is dimensionally mapped through the fully connected layer and the GeLU activation function, mapping the feature vector to a low-dimensional space, abstractly representing the features, highlighting the key information in the features, and enabling the model to learn the complex nonlinear relationship between features. , represents the fully connected layer, Represents the GeLU activation function. The feature vector after dimension mapping passes through the fully connected layer and the Sigmoid function to generate the spatial dimension attention factor , channel dimension attention factor and edge dimension attention factor , , represents the fully connected layer, Represents the Sigmoid function. A gating mechanism is introduced after the three attention factors to dynamically adjust the activation degree of the weights through control signals, thereby improving the ability of the attention factors to select key features. , , , represents the gating mechanism, , and They represent the dynamically adjusted spatial dimension attention factor, channel dimension attention factor, and edge dimension attention factor, respectively. The attention factor is used to adjust the traditional convolution Perform dynamic adjustments to obtain dynamic convolution , Represents element-wise multiplication, and finally uses dynamic convolution Features Perform convolution operations to accurately extract local key information from input features while suppressing the influence of irrelevant information. , Represents a dynamic convolution operation.

[0013] Preferably, in step S2, the dimension of the input feature is reduced by average pooling to reduce the amount of calculation and parameter complexity, and then the feature is mapped to a low-dimensional space using a fully connected layer and a GeLU activation function to extract key information. Next, a multi-dimensional attention factor is generated through a fully connected layer and a Sigmoid, and these attention factors are dynamically adjusted using a gating mechanism to enhance their ability to select key information. Finally, the adjusted attention factor is used for traditional convolution to generate dynamic convolution to achieve accurate convolution of the input feature, which not only extracts local key information, but also suppresses the influence of irrelevant information, thereby improving the accuracy and efficiency of local feature extraction.

[0014] Preferably, in step S2, according to the variance of the input image Dynamically adjust the weights of the fully connected layer ,set up To adjust the coefficients, increase the scaling of the weights when the variance is large: ,in is the variance threshold, which is used to determine whether the degradation is serious.

[0015] Preferably, in step S2, dynamically adjusting the weights of the fully connected layer can dynamically adjust the parameters of the model according to the quality of the input image, so that the model has better robustness and adaptability when processing changes in image quality, thereby improving the effectiveness of feature extraction.

[0016] Preferably, in step S3, for the dynamic local attention module, the input image feature First, the input features Perform batch normalization to ensure the stability of the distribution of feature values ​​and improve training stability. Extract local features through 1×1 convolutional layers and 3×3 convolutional layers to capture contextual information within the perception field of view, and then enhance important features through local self-attention. , represents the batch normalization operation, represents a 1×1 convolutional layer, represents a 3×3 convolutional layer, Represents local self-attention, effectively aggregates local context information through adaptive convolution blocks, then uses 1×1 convolution layers to compress features, and finally introduces residual connections to alleviate the gradient vanishing problem , represents a 1×1 convolutional layer, represents the adaptive convolution block, Represents addition of elements.

[0017] Preferably, in step S3, the goal is to stabilize the feature distribution by batch normalizing the input image features, and then use 1×1 and 3×3 convolutions to extract local context information. Then, important features are strengthened through the local self-attention mechanism to improve the feature expression ability of the local area. On this basis, the local features are effectively aggregated using adaptive convolution blocks, and finally the features are compressed using 1×1 convolutions, and the effective transmission of features is maintained through residual connections to alleviate the gradient vanishing problem. The role of the entire module is to enhance the expression ability of local features through dynamic local attention and adaptive aggregation, ensuring that key information is retained during the convolution process, thereby improving the overall performance and stability of the model.

[0018] Preferably, in step S4, for the global context feature extraction module, the image feature is input After the layer normalization operation, the features of each channel are normalized to make them more stable in mean and variance, which helps to accelerate model training and improve the generalization ability of the model. Then, the dilated convolution is used to increase the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation. , Representation layer normalization operation, represents the dilated convolution, k represents the convolution kernel size, d represents the dilation rate, and then Q, K and V are obtained through a 1×1 convolution layer and a depth 3×3 convolution layer. , , , represents a 1×1 convolutional layer, represents a 3×3 convolutional layer with depth. Q, K, and V represent query, key, and value, respectively. In order to adjust the shapes of Q, K, and V to a form suitable for matrix multiplication, Q, K, and V need to be reshaped. , , , Represents a reshaping operation. and Perform matrix multiplication and then calculate the attention map through the Softmax function to indicate the degree of correlation between different positions. Perform weighted summation according to the attention map to obtain the feature representation weighted by attention , represents the Softmax function, represents element-wise multiplication, It represents element addition, which is then transformed through a 1×1 convolutional layer, and then a residual connection is introduced to alleviate the gradient vanishing problem. , represents a 1×1 convolutional layer, It represents element addition, and then uses 1×1 convolution layer and ReLU activation function to expand the number of feature channels and increase nonlinearity. A 7×7 convolution layer is used to extract global spatial information. Then a 1×1 convolution layer is used to restore the channel dimension. Finally, a residual connection is introduced to alleviate the gradient disappearance problem. , represents a 1×1 convolutional layer, represents the ReLU activation function, Represents addition of elements.

[0019] Preferably, in step S4, the stability and receptive field of the features are improved through layer normalization and dilated convolution operations, so as to better capture the global context information of the input image. Q, K and V are calculated through 1×1 convolution layers and deep 3×3 convolution layers, and the correlation between different positions is calculated using the attention mechanism to obtain weighted feature representation, which strengthens the model's understanding of global information. Subsequently, residual connections, convolutions and ReLU activation operations are introduced to ensure the transmission of features, alleviate the gradient vanishing problem, and further expand the feature channel to enhance nonlinear relationships. The whole process aims to effectively extract global information in the image and enhance the model's ability to accurately represent and generalize features.

[0020] Preferably, in step S5, the multidimensional feature enhancement network is composed of an encoder, a global context feature extraction module, and a decoder, wherein the encoder is composed of a dynamic local attention module and downsampling. In the encoder, the dynamic local attention module is first used to extract local features, and then the spatial resolution is reduced by downsampling to aggregate more local context information. , represents the dynamic local attention module, represents downsampling, and are the two outputs of the encoder, Input to the decoder for feature fusion, As the input of the next encoder, the decoder consists of upsampling and dynamic local attention modules. In the decoder, the features are gradually restored to the original resolution by upsampling, and then the upsampled features are compared with the features of the encoder input. Perform feature fusion, and use the dynamic local attention module to optimize the features and enhance the image quality. , represents the dynamic local attention module, represents upsampling, As the input of the next decoder, the shallow features extracted by the shallow feature extraction module are input , use the encoder to extract local multi-scale features, then use the global context feature extraction module to extract global features, and finally use the decoder to gradually restore the features to the resolution at the input, and combine the information of the encoder jump connection to obtain enhanced image features.

[0021] Preferably, in step S5, the multidimensional feature enhancement network includes an encoder, a global context feature extraction module and a decoder. The encoder extracts local features through a dynamic local attention module and uses downsampling to reduce the spatial resolution to aggregate more context information. The decoder restores the original resolution of the features through upsampling, and fuses the upsampled features with the output features of the encoder, and then further optimizes the features through the dynamic local attention module to enhance the image quality. The entire network combines the jump connections of the encoder and decoder, first extracts multi-scale local features, then fuses the global features, and finally gradually restores to the input resolution, thereby obtaining high-quality enhanced image features.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The multi-dimensional feature enhancement network proposed in the present invention combines CNN and Transformer to achieve efficient image enhancement. By using CNN modules to process the encoder and decoder and introducing the Transformer module in the global context feature extraction module, the advantages of CNN in local feature modeling and the ability of Transformer in global feature extraction are fully utilized. In addition, by designing adaptive convolution blocks to capture multiple attention mechanisms, the local feature extraction effect is further enhanced, thereby achieving a good balance between the performance and computational complexity of image enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention provides a flow chart of the method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization.

[0024] Figure 2 This is a structural diagram of the adaptive convolution block provided by the present invention.

[0025] Figure 3 This is a structural diagram of the dynamic local attention module provided by the present invention.

[0026] Figure 4 It is a structural diagram of the global context feature extraction module provided by the present invention.

[0027] Figure 5 This is a multi-dimensional feature enhancement network structure diagram provided by the present invention.

[0028] Figure 6 It is a comparison diagram of the ultrasonic image of the rectus abdominis muscle before and after enhancement for postpartum rehabilitation optimization provided by the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the accompanying drawings of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] See also Figures 1 to 6 The present invention provides a rectus abdominis ultrasound image enhancement method for postpartum rehabilitation optimization. The encoder and decoder are processed by using a CNN module, and a Transformer block is introduced in the global context feature extraction module, which fully utilizes the advantages of CNN in local feature modeling and the ability of Transformer in global feature extraction. In addition, the local feature extraction effect is further enhanced by designing an adaptive convolution block to capture multiple attention mechanisms.

[0031] See also Figure 1 As shown, a rectus abdominis ultrasound image enhancement method for optimizing postpartum rehabilitation in an embodiment of the present application.

[0032] S1. Use a 5-10MHz high-frequency probe to obtain ultrasound images of the rectus abdominis muscle for postpartum rehabilitation optimization. Obtain 1,000 ultrasound images of the rectus abdominis muscle and annotate the acquired ultrasound images of the rectus abdominis muscle. The annotated images constitute a dataset of ultrasound images of the rectus abdominis muscle. Divide the dataset into a training set and a validation set in a ratio of 7:3.

[0033] S2. Construct an adaptive convolution block, including an average pooling layer, a fully connected layer, a GeLU activation function, and a traditional convolution, learn the attention of the three dimensions of space, channel, and edge in the image features, reduce the input feature dimension through average pooling, reduce the amount of calculation and parameter complexity, and then use the fully connected layer and GeLU activation to extract key information, generate a multi-dimensional attention factor and dynamically adjust it through a gating mechanism to enhance its selectivity, and finally combine the adjusted attention factor to generate a dynamic convolution, accurately extract local features and suppress irrelevant information, and improve the accuracy and efficiency of local feature extraction.

[0034] Furthermore, if Figure 2 As shown, the input image features , using an average pooling layer to reduce Dimension, reducing the parameters and computational complexity of the model. , Represents the average pooling operation. The pooling result is dimensionally mapped through the fully connected layer and the GeLU activation function, mapping the feature vector to a low-dimensional space, abstractly representing the features, highlighting the key information in the features, and enabling the model to learn the complex nonlinear relationship between features. , represents the fully connected layer, Represents the GeLU activation function. The feature vector after dimension mapping passes through the fully connected layer and the Sigmoid function to generate the spatial dimension attention factor , channel dimension attention factor and edge dimension attention factor , , represents the fully connected layer, Represents the Sigmoid function. A gating mechanism is introduced after the three attention factors to dynamically adjust the activation degree of the weights through control signals, thereby improving the ability of the attention factors to select key features. , , , represents the gating mechanism, , and They represent the dynamically adjusted spatial dimension attention factor, channel dimension attention factor, and edge dimension attention factor, respectively. The attention factor is used to adjust the traditional convolution Perform dynamic adjustments to obtain dynamic convolution , Represents element-wise multiplication, and finally uses dynamic convolution Features Perform convolution operations to accurately extract local key information from input features while suppressing the influence of irrelevant information. , Represents a dynamic convolution operation.

[0035] Furthermore, in step S2, according to the variance of the input image Dynamically adjust the weights of the fully connected layer , set the adjustment coefficient to 0.5, the variance threshold to 0.8, and increase the weight scaling ratio when the variance is large: .

[0036] S3. Construct a dynamic local attention module, including batch normalization layer, local self-attention and adaptive convolution blocks, to extract local information and enhance feature expression. Batch normalization is used to stabilize feature distribution, 1×1 and 3×3 convolutions are used to extract contextual information, and local self-attention mechanism is used to enhance important features. The adaptive convolution block aggregates local features, and then compresses features through 1×1 convolution and adds residual connection to ensure effective transmission and alleviate gradient vanishing, ultimately improving the expressiveness of local features and the stability of the model.

[0037] Furthermore, if Figure 3 As shown, the input image features First, the input features Perform batch normalization to ensure the stability of the distribution of feature values ​​and improve training stability. Extract local features through 1×1 convolutional layers and 3×3 convolutional layers to capture contextual information within the perception field of view, and then enhance important features through local self-attention. , represents the batch normalization operation, represents a 1×1 convolutional layer, represents a 3×3 convolutional layer, Represents local self-attention, effectively aggregates local context information through adaptive convolution blocks, then uses 1×1 convolution layers to compress features, and finally introduces residual connections to alleviate the gradient vanishing problem , represents a 1×1 convolutional layer, represents the adaptive convolution block, Represents addition of elements.

[0038] S4. Construct a global context feature extraction module, including layer normalization, dilated convolution, deep convolution layer and Softmax function. Improve the stability and receptive field of features through layer normalization and dilated convolution. Combined with the attention mechanism, calculate the position correlation and obtain weighted feature representation, strengthen the model's understanding of global information, introduce residual connection, convolution and ReLU activation to ensure feature transmission and expansion channel, alleviate gradient disappearance, so as to effectively extract global information, improve the accuracy of feature representation and the generalization ability of the model.

[0039] Furthermore, if Figure 4 As shown, the input image features After the layer normalization operation, the features of each channel are normalized to make them more stable in mean and variance, which helps to accelerate model training and improve the generalization ability of the model. Then, the dilated convolution is used to increase the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation. , Representation layer normalization operation, represents the dilated convolution, k represents the convolution kernel size, d represents the dilation rate, and then Q, K and V are obtained through a 1×1 convolution layer and a depth 3×3 convolution layer. , , , represents a 1×1 convolutional layer, represents a 3×3 convolutional layer with depth. Q, K, and V represent query, key, and value, respectively. In order to adjust the shapes of Q, K, and V to a form suitable for matrix multiplication, Q, K, and V need to be reshaped. , , , Represents a reshaping operation. and Perform matrix multiplication and then calculate the attention map through the Softmax function to indicate the degree of correlation between different positions. Perform weighted summation based on the attention map to obtain the feature representation weighted by attention , represents the Softmax function, represents element-wise multiplication, It represents element addition, which is then transformed through a 1×1 convolutional layer, and then a residual connection is introduced to alleviate the gradient vanishing problem. , represents a 1×1 convolutional layer, It represents element addition, and then uses 1×1 convolution layer and ReLU activation function to expand the number of feature channels and increase nonlinearity. A 7×7 convolution layer is used to extract global spatial information. Then a 1×1 convolution layer is used to restore the channel dimension. Finally, a residual connection is introduced to alleviate the gradient disappearance problem. , represents a 1×1 convolutional layer, represents the ReLU activation function, Represents addition of elements.

[0040] S5. Construct a multi-dimensional feature enhancement network, which consists of an encoder, a global context feature extraction module, and a decoder. The encoder consists of a dynamic local attention module and downsampling, and the decoder consists of an upsampling and dynamic local attention module.

[0041] Furthermore, if Figure 5 As shown in the figure, the multi-dimensional feature enhancement network consists of an encoder, a global context feature extraction module, and a decoder. Three encoders and three decoders are used to extract local features. The encoder consists of a dynamic local attention module and downsampling. In the encoder, the dynamic local attention module is first used to extract local features, and then the spatial resolution is reduced by downsampling to aggregate more local context information. , represents the dynamic local attention module, represents downsampling, and are the two outputs of the encoder, Input to the decoder for feature fusion, As the input of the next encoder, the decoder consists of upsampling and dynamic local attention modules. In the decoder, the features are gradually restored to the original resolution by upsampling, and then the upsampled features are compared with the features of the encoder input. Perform feature fusion, and use the dynamic local attention module to optimize the features and enhance the image quality. , represents the dynamic local attention module, represents upsampling, As the input of the next decoder, the shallow features extracted by the shallow feature extraction module are input To the encoder, the encoder is used to gradually extract local multi-scale features, and then the global context feature extraction module is used to extract global features. Finally, the decoder is used to gradually restore the features to the resolution at the input, and the information of the encoder jump connection is combined to obtain the enhanced image features.

[0042] S6. Construct a rectus abdominis ultrasound image enhancement model, which consists of a shallow feature extraction module, a multidimensional feature recovery network and an image reconstruction module. Input the rectus abdominis ultrasound image optimized for postpartum rehabilitation into the model to obtain an enhanced rectus abdominis ultrasound image optimized for postpartum rehabilitation.

[0043] Further, in step S6, a rectus abdominis ultrasound image enhancement model is developed using the Python language through the Pycharm application and implemented based on the Pytorch framework. The model inputs the rectus abdominis ultrasound image optimized for postpartum rehabilitation and starts training using a self-supervised learning method. The model effectively extracts local and global features from the image and obtains the enhanced rectus abdominis ultrasound image optimized for postpartum rehabilitation through an image reconstruction module.

[0044] Furthermore, if Figure 6 As shown, Figure 6 The middle left part is the original ultrasound image of the rectus abdominis muscle. Figure 6 The middle right half is the enhanced rectus abdominis ultrasound image obtained after being processed by the rectus abdominis ultrasound image enhancement model, which significantly enhances the clarity and recognizability of the rectus abdominis ultrasound image and improves the accuracy of subsequent detection and processing.

[0045] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization, characterized in that: The following steps are involved: S1. Obtain the image to be processed, and use a high-frequency probe to obtain an ultrasound image of the rectus abdominis muscle for postpartum rehabilitation optimization; S2, build an adaptive convolution block, including average pooling layer, fully connected layer, GeLU activation function and traditional convolution, learn the attention of three dimensions of space, channel and edge in image features, and effectively aggregate local context information; S3, build a dynamic local attention module, including batch normalization layer, local self-attention and adaptive convolution blocks, to extract local information; S4, construct a global context feature extraction module, including layer normalization, hole convolution, deep convolution layer and Softmax function, to extract global context information; S5. Construct a multi-dimensional feature enhancement network, which consists of an encoder, a global context feature extraction module and a decoder. The encoder consists of a dynamic local attention module and downsampling, and the decoder consists of an upsampling and dynamic local attention module. S6. Construct a rectus abdominis ultrasound image enhancement model, which consists of a shallow feature extraction module, a multidimensional feature enhancement network and an image reconstruction module. Input the rectus abdominis ultrasound image optimized for postpartum rehabilitation into the model to obtain an enhanced rectus abdominis ultrasound image optimized for postpartum rehabilitation.

2. The method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization according to claim 1, characterized in that: In step S2, for the adaptive convolution block, the input image features , using an average pooling layer to reduce Dimension, reducing the parameters and computational complexity of the model. , Represents the average pooling operation. The pooling result is dimensionally mapped through the fully connected layer and the GeLU activation function, mapping the feature vector to a low-dimensional space, abstractly representing the features, highlighting the key information in the features, and enabling the model to learn the complex nonlinear relationship between features. , represents the fully connected layer, Represents the GeLU activation function. The feature vector after dimension mapping passes through the fully connected layer and the Sigmoid function to generate the spatial dimension attention factor , channel dimension attention factor and edge dimension attention factor , , represents the fully connected layer, Represents the Sigmoid function. A gating mechanism is introduced after the three attention factors to dynamically adjust the activation degree of the weights through control signals, thereby improving the ability of the attention factors to select key features. , , , represents the gating mechanism, , and They represent the dynamically adjusted spatial dimension attention factor, channel dimension attention factor, and edge dimension attention factor, respectively. The attention factor is used to adjust the traditional convolution Perform dynamic adjustments to obtain dynamic convolution , Represents element-wise multiplication, and finally uses dynamic convolution Features Perform convolution operations to accurately extract local key information from input features while suppressing the influence of irrelevant information. , Represents a dynamic convolution operation.

3. The method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization according to claim 2, characterized in that: In the adaptive convolution block, according to the variance of the input image Dynamically adjust the weights of the fully connected layer ,set up To adjust the coefficients, increase the scaling of the weights when the variance is large: ,in is the variance threshold, which is used to determine whether the degradation is serious.

4. The method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization according to claim 1, characterized in that: In step S3, for the dynamic local attention module, the input image features , firstly, the input features Perform batch normalization to ensure the stability of the distribution of feature values ​​and improve training stability. Extract local features through 1×1 convolutional layers and 3×3 convolutional layers to capture contextual information within the perception field of view, and then enhance important features through local self-attention. , represents the batch normalization operation, represents a 1×1 convolutional layer, represents a 3×3 convolutional layer, Represents local self-attention, effectively aggregates local context information through adaptive convolution blocks, then uses 1×1 convolution layers to compress features, and finally introduces residual connections to alleviate the gradient vanishing problem , represents a 1×1 convolutional layer, represents the adaptive convolution block, Represents addition of elements.

5. The method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization according to claim 1, characterized in that: In step S4, for the global context feature extraction module, the input image features After the layer normalization operation, the features of each channel are normalized to make them more stable in mean and variance, which helps to accelerate model training and improve the generalization ability of the model. Then, the dilated convolution is used to increase the receptive field of the convolution kernel without increasing the number of parameters and the amount of calculation. , Representation layer normalization operation, represents the dilated convolution, k represents the convolution kernel size, d represents the dilation rate, and then Q, K and V are obtained through a 1×1 convolution layer and a depth 3×3 convolution layer. , , , represents a 1×1 convolutional layer, represents a 3×3 convolutional layer with depth. Q, K, and V represent query, key, and value, respectively. In order to adjust the shapes of Q, K, and V to a form suitable for matrix multiplication, Q, K, and V need to be reshaped. , , , Represents a reshaping operation. and Perform matrix multiplication and then calculate the attention map through the Softmax function to indicate the degree of correlation between different positions. Perform weighted summation according to the attention map to obtain the feature representation weighted by attention , represents the Softmax function, represents element-wise multiplication, It represents element addition, which is then transformed through a 1×1 convolutional layer, and then a residual connection is introduced to alleviate the gradient vanishing problem. , represents a 1×1 convolutional layer, It represents element addition, and then uses 1×1 convolution layer and ReLU activation function to expand the number of feature channels and increase nonlinearity. A 7×7 convolution layer is used to extract global spatial information. Then a 1×1 convolution layer is used to restore the channel dimension. Finally, a residual connection is introduced to alleviate the gradient disappearance problem. , represents a 1×1 convolutional layer, represents a 7×7 convolutional layer, represents the ReLU activation function, Represents addition of elements.

6. The method for enhancing the ultrasonic image of the rectus abdominis muscle for postpartum rehabilitation optimization according to claim 1, characterized in that: In step S5, The multi-dimensional feature enhancement network consists of an encoder, a global context feature extraction module, and a decoder. The encoder consists of a dynamic local attention module and downsampling. In the encoder, the dynamic local attention module is first used to extract local features, and then downsampling is used to reduce the spatial resolution and aggregate more local context information. , represents the dynamic local attention module, represents downsampling, and are the two outputs of the encoder, Input to the decoder for feature fusion, As the input of the next encoder, the decoder consists of upsampling and dynamic local attention modules. In the decoder, the features are gradually restored to the original resolution by upsampling, and then the upsampled features are compared with the features of the encoder input. Perform feature fusion, and use the dynamic local attention module to optimize the fused features and enhance the image quality. , represents the dynamic local attention module, represents upsampling, As the input of the next decoder, the shallow features extracted by the shallow feature extraction module are input To the encoder, the encoder is used to extract local multi-scale features, and then the global context feature extraction module is used to extract global features. Finally, the decoder is used to gradually restore the features to the resolution at the input, and the information of the encoder jump connection is combined to obtain the enhanced image features.

Images