Retina image segmentation method based on lightweight two-way cascade network
By adopting a lightweight dual-path cascade network in the retinal image segmentation model, combining structured pruning, asymmetric convolution and coordinate attention modules, the problems of vascular discontinuity and large parameters are solved, and a more efficient retinal Image segmentation effect.
Patent Information
- Application Number
- CN202411131172.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-17
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing retinal image segmentation model is prone to vascular discontinuity when processing fundus images, and the complex network structure leads to large parameters and heavy calculation load, which limits the application and scalability of the algorithm.
Using a lightweight dual-path cascade network method, the improved U-shaped network is introduced, combined with structured pruning, asymmetric convolution and coordinate attention modules, the model parameter quantity and calculation cost are reduced, and the dual-path information is used. The fusion module performs cascade processing to enhance feature extraction capabilities.
It realizes more accurate retinal vascular segmentation, reduces model complexity and calculation cost, improves segmentation accuracy and edge recognition capabilities, and is suitable for actual environments.
Smart Images

Figure CN119067991A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and more specifically to a retinal image segmentation method based on a lightweight dual-path cascade network. Background Art
[0002] The vascular structure and morphology of fundus images contain rich information. Retinal image segmentation plays a vital role in assisting the screening of eye diseases. However, due to the low contrast between the tiny blood vessels and the background in fundus images and the complex tree-like branching structure of blood vessels, accurate retinal image segmentation is still a challenging task.
[0003] Existing models are prone to vascular discontinuity in segmentation results, and many current studies focus on improving segmentation performance while paying less attention to model efficiency. Complex network structures often bring a large number of parameters and computational loads, which not only consumes a lot of computing resources, but also limits the application and scalability of the algorithm in practical environments.
[0004] Therefore, how to solve the problem of discontinuous blood vessel segmentation and large number of model parameters in the fundus image segmentation task is an urgent problem that technicians in this field need to solve. Summary of the invention
[0005] In view of this, the present invention provides a retinal image segmentation method based on a lightweight dual-path cascade network to solve the problems of discontinuous blood vessel segmentation and large number of model parameters in fundus image segmentation tasks.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides a retinal image segmentation method based on a lightweight two-way cascade network, comprising the following steps:
[0008] Two U-shaped networks with the same structure are introduced and improved as the front-end network and the back-end network respectively. The front-end network and the back-end network are cascaded to construct a retinal image segmentation model.
[0009] In the previous network, the encoder part is processed through two stages. Each stage contains a residual module and a maximum pooling layer to downsample the original input retinal image. After one stage of processing, the resolution of the feature map is halved and the number of feature channels is doubled. In each stage, the feature map output by the residual module has two flow paths: one path continues to downsample to extract deeper semantic information; the other path enters the skip path and is connected to the decoder part after being processed by the coordinate attention module.
[0010] After the encoder part is completed, the feature map is processed by a residual module and then the decoding process begins;
[0011] The decoding part consists of two stages. In each stage, upsampling is performed through deconvolution and residual modules to restore the size of the feature map and output the first feature map through 1×1 convolution.
[0012] The first feature map output by the previous network is superimposed with the original input retinal image as the input of the subsequent network;
[0013] The subsequent network processes the first feature map according to the same encoding and decoding process as the previous network to obtain the final retinal vascular image.
[0014] Furthermore, the final retinal vascular image is expressed as:
[0015] U(x)=U (2) (x,U (1) (x)
[0016] Among them, U(x) represents the final retinal vascular image, x represents the original input retinal image, and U (1) (x) represents the first feature map output by the previous network, U (1) represents the previous network, U (2) Represents the subsequent network.
[0017] Further improvements to the U-type network include: using a structured pruning method to prune the five-layer architecture of the original U-type network to three layers, and pruning the feature channels of the original U-type network to {32, 64, 128}.
[0018] Furthermore, the processing process of the coordinate attention module for the feature map of size W×H includes:
[0019] Using pooling kernels of size (H, 1) and (1, W), the feature map of each channel is encoded along the x-axis and y-axis respectively to obtain the direction-aware feature map. The calculation formula is:
[0020]
[0021]
[0022] in, represents the output of the cth channel with height h, and the size is C×H×1, represents the output of the cth channel with a width of w, and the size is C×1×W; x(i,h) represents the value of the feature map with a width coordinate of i and a height of h along the cth channel in the x-axis direction, and x(w,j) represents the value of the feature map with a width of w and a height coordinate of j along the y-axis direction;
[0023] Direction-aware feature maps via coordinate attention Combined, the intermediate feature representation is generated through 1×1 convolution processing, and the calculation formula is:
[0024]
[0025] Where · represents the connection operation in the spatial dimension, δ represents the nonlinear activation function, F1 represents the 1x1 convolution transformation function, and f∈R C / r*(H+w) Represents the intermediate feature map that encodes vertical and horizontal information;
[0026] The intermediate feature representation f is decomposed into two directions: width and height, respectively: f h ∈R C / r*H and f w ∈R C / r*w , respectively, after 1×1 convolution F h 、F w And Sigmoid activation function δ is processed to match the number of channels of the input feature map to obtain g h and g w , the process is as follows:
[0027] g h =δ(F h (f h ))
[0028] g w =δ(F w (f w ))
[0029] The calculated weight g h and g w Applied to the feature map x(i,j) after downsampling by the residual module and the maximum pooling layer, the final weighted feature representation Y=[y1,y2,…,y t ], so that the retinal image segmentation model focuses on the vascular area during training, where
[0030] Furthermore, a dual-path information fusion module is introduced to cascade the previous network and the next network, and the first feature map output by the previous network is superimposed with the original retinal feature map.
[0031] Furthermore, the step of superimposing the first feature map output by the previous network with the original retinal feature map by the dual-path information fusion module includes:
[0032] The original input retinal image of the previous network is subjected to multi-scale feature extraction, and then subjected to one 1×1 convolution, three 3×3 dilated convolutions, and one global average pooling operation. After concatenation, the first output feature map is obtained through a 1×1 convolution.
[0033] Perform multi-scale feature extraction on the first feature map output by the previous network, and sequentially undergo a 1×1 convolution, three 3×3 dilated convolutions, and a global average pooling operation, and then concatenate and pass a 1×1 convolution to obtain the second output feature map;
[0034] The first output feature map is concatenated with the second output feature map in the channel dimension as the input of the subsequent network.
[0035] Further improvements to the U-type network include: introducing asymmetric convolution.
[0036] In a second aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; the characterised in that when the processor executes the computer program, the steps of the above-mentioned retinal image segmentation method based on a lightweight dual-path cascade network are implemented.
[0037] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the above-mentioned retinal image segmentation method based on a lightweight two-way cascade network are implemented.
[0038] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:
[0039] First of all, taking into account the problems of semantic information stacking and dilution in deep multi-channel structures of existing network structures, as well as high computational costs, the present invention adopts structured pruning to reduce the number of network layers and feature channels, and introduces asymmetric convolution to further reduce the number of parameters and computational complexity of the model. The standard convolution block is replaced by a more efficient residual module to promote information flow and reduce excessive stacking of semantic information, thereby constructing a lightweight network architecture.
[0040] Secondly, in order to fully integrate deep semantic information with shallow detail information and improve the model's classification accuracy and edge recognition ability for objects in images, the present invention introduces a coordinate attention module into the jump connection in the codec structure. The introduction of this module can enhance the network's ability to capture important vascular features and reduce the impact of noise information, so that the network can use detail information more accurately.
[0041] Finally, the present invention cascades two improved U-type networks by adopting a dual-path information fusion module. This cascaded network structure provides a secondary learning opportunity for the feature map and further refines the feature representation. The U-type network in the first stage is defined as the "front-stage network", and the U-type network in the second stage is defined as the "back-stage network". This dual-path cascade strategy not only enhances the feature extraction capability, but also provides a more effective solution for complex retinal image segmentation tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0043] Figure 1 A schematic diagram of the structure of the retinal image segmentation model provided by the present invention;
[0044] Figure 2 A schematic diagram of the structure of a U-shaped network before pruning provided by the present invention;
[0045] Figure 3 A schematic diagram of the structure of a U-shaped network after pruning provided by the present invention;
[0046] Figure 4 A schematic diagram of asymmetric convolution provided by the present invention;
[0047] Figure 5 This is a structural schematic diagram of the dual-path information fusion module provided by the present invention. DETAILED DESCRIPTION
[0048] The following will be combined with the drawings in the embodiments 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.
[0049] like Figure 1 As shown, the embodiment of the present invention discloses a retinal image segmentation method based on a lightweight two-way cascade network, comprising the following steps:
[0050] Two U-shaped networks with the same structure are introduced and improved as the front-end network and the back-end network respectively. The front-end network and the back-end network are cascaded to construct a retinal image segmentation model.
[0051] In the previous network, the encoder part is processed through two stages. Each stage contains a residual module and a maximum pooling layer to downsample the original input retinal image. After one stage of processing, the resolution of the feature map is halved and the number of feature channels is doubled. In each stage, the feature map output by the residual module has two flow paths: one path continues to downsample to extract deeper semantic information; the other path enters the skip path and is connected to the decoder part after being processed by the coordinate attention module (CA).
[0052] After the encoder part is completed, the feature map is processed by a residual module and then the decoding process begins;
[0053] The decoding part consists of two stages. In each stage, upsampling is performed through deconvolution and residual modules to restore the size of the feature map and output the first feature map through 1×1 convolution.
[0054] The first feature map output by the previous network is superimposed with the original input retinal image as the input of the subsequent network;
[0055] The subsequent network processes the first feature map according to the same encoding and decoding process as the previous network to obtain the final retinal vascular image.
[0056] The embodiment of the present invention aims to convert the original input retinal image x into an attention-weighted feature map to compensate for the features that may be lost in the previous network. (2) Follow the same principles as the previous network U (1) The same encoding and decoding process is used to further refine the feature map and output a refined segmentation result.
[0057] The final retinal vascular image is represented as:
[0058] U(x)=U (2) (x,U (1) (x)
[0059] Among them, U(x) represents the final retinal vascular image, x represents the original input retinal image, and U (1) (x) represents the first feature map output by the previous network, U (1) represents the previous network, U (2) Represents the subsequent network.
[0060] The present invention improves the network architecture in the following aspects.
[0061] (1) Structured pruning.
[0062] There are two main problems in the development of retinal image segmentation algorithms. First, too many feature channels will make the model too complex, which not only increases training time and storage space requirements, but also may introduce redundant features, which does not substantially help improve model accuracy. These similar features cause unnecessary overlaps in different channels and may cause the model to overfit when dealing with binary classification tasks, affecting generalization capabilities. Secondly, although the pooling operation can reduce the resolution of the feature map and increase the receptive field, it may cause the loss of important information during the downsampling process. In particular, edge and texture information, once lost, is difficult to recover in the subsequent upsampling process. The details of tiny blood vessels are particularly easy to be lost in this process. Excessive pooling will dilute the vascular information and reduce segmentation accuracy.
[0063] The present invention draws on the strategy of structured pruning to design a lightweight network, with the goal of finding a suitable retinal segmentation network structure, that is, optimizing the network by reducing redundant channels and layers. Specifically, the number of feature channels is reduced to reduce network complexity, and the number of pooling layers is reduced to retain more boundary and context information. In this simplification process, the five-layer architecture of the original U-type network is pruned to three layers, and the number of feature channels is reduced accordingly, and the feature channels of the original U-type network are pruned from {64, 128, 256, 512, 1024} to {32, 64, 128}.
[0064] like Figure 2-Figure 3 The figure shows the comparison between the original U-Net structure and the U-Net structure after lightweight processing. The blue blocks are convolution or deconvolution layers, and the black numbers represent the number of feature channels. Although the number of network parameters after lightweight processing is only 0.7M, it still retains enough feature channels to capture important information about the vascular structure, achieving an effective balance between segmentation performance and model size.
[0065] (2) Asymmetric convolution optimization
[0066] Improvements to the U-type network also include: asymmetric convolution optimization of the convolution kernel, the process of which is: decomposing the n×n convolution kernel into heterogeneous n×1 and 1×n convolution kernel sequences.
[0067] In order to balance the performance and efficiency of the model and ensure that the complexity remains within a reasonable range, this embodiment uses a convolution kernel of size 3×3. Figure 4 As shown in Figure 5, by replacing all standard 3×3 convolutions with a series of 3×1 and 1×3 convolutions, this replacement strategy significantly reduces the computational cost and number of parameters of the model.
[0068] From a mathematical point of view, the total number of parameters of a 3×3 convolution kernel is 9C 2 In contrast, heterogeneous convolution only requires 6C 2If these asymmetric convolutions maintained the same number of channels as the original square convolution kernel, the number of parameters would be reduced by one third.
[0069] (3) Coordinate attention module.
[0070] In fundus retinal imaging, distinguishing blood vessels from background is a major challenge because blood vessel pixels are relatively scarce compared to the background. In order to enable the model to pay attention to the vascular area and strengthen the training of the vascular area, it is a crucial issue. The present invention adopts a coordinate attention mechanism, which aggregates vascular features along two orthogonal axes, horizontal and vertical. During the training process, the dependencies between blood vessels and the precise spatial information of their positions can be captured, and finally an attention matrix is formed. Subsequently, this attention matrix is combined with the feature map to effectively enhance the representation of the vascular area in the feature map.
[0071] The processing process of the coordinate attention module on the feature map includes:
[0072] The coordinate attention mechanism exploits the fact that global pooling operations are widely used to encode spatial information, but these operations often fall short in preserving position information, which is crucial for visual tasks. To overcome this limitation, when processing feature maps of size W×H, pooling kernels of size (H, 1) and (1, W) are used to encode the feature maps of each channel along the x-axis and y-axis, respectively, to obtain direction-aware feature maps, calculated as:
[0073]
[0074] The above transformation is performed along the x-axis and y-axis, where represents the output of the cth channel with height h, and the size is C×H×1, represents the output of the cth channel with a width of w and a size of C×1×W; x(i,h) represents the value of the feature map with a width coordinate of i and a height of h along the cth channel in the x-axis direction, and x(w,j) represents the value of the feature map with a width of w and a height coordinate of j along the y-axis direction; compared to ordinary channel attention, coordinate attention can obtain attention weights in two directions, so that the network can focus more on the area of interest.
[0075] Direction-aware feature maps via coordinate attention Combined, the intermediate feature representation is generated through 1×1 convolution processing, and the calculation formula is:
[0076] f∈δ(F1(Z h ·Z w ))
[0077] Where · represents the connection operation in the spatial dimension, δ represents the nonlinear activation function, F1 represents the 1x1 convolution transformation function, and f∈R C / r*(H+w) Represents the intermediate feature map that encodes vertical and horizontal information;
[0078] The intermediate feature representation f is decomposed into two directions: width and height, respectively: f h ∈R C / r*H and f w ∈R C / r*w , respectively, after 1×1 convolution F h 、F w And Sigmoid activation function δ is processed to match the number of channels of the input feature map to obtain g h and g w , the process is as follows:
[0079] g h =δ(F h (f h ))
[0080] g w =δ(F w (f w ))
[0081] After this processing, the attention weights g in two directions are obtained h and g w The calculated weights are applied to the feature map x(i,j) after downsampling by the residual module and the maximum pooling layer to obtain the final weighted feature representation Y=[y1,y2,…,y t ], so that the retinal image segmentation model focuses on the vascular area during training, where
[0082] (4) Dual-path information fusion module.
[0083] Although the atrous convolution expands the receptive field by skipping input pixels, effectively improving the network's ability to extract information, this method may lose pixel information and produce the so-called "grid effect". In the detailed retinal image segmentation task, this effect may cause the network to ignore those vascular pixels that are not directly covered by the convolution kernel. In order to address this problem, the present invention does not directly replace the traditional convolution layer with the atrous convolution layer, but introduces a dual-path information fusion module (DPFM) at the interface of the cascade network, cascades the previous network and the next network, and superimposes the first feature map output by the previous network with the original retinal feature map as the input of the next network, providing the next network with richer multi-scale vascular information.
[0084] like Figure 5As shown in FIG. 1 , it is a schematic diagram of the specific structure of the dual-path information fusion module, and the step of superimposing the first feature map output by the previous network with the original retinal feature map includes:
[0085] The original input retinal image of the previous network is subjected to multi-scale feature extraction, and then subjected to one 1×1 convolution, three 3×3 dilated convolutions, and one global average pooling operation. After concatenation, the first output feature map is obtained through a 1×1 convolution.
[0086] Perform multi-scale feature extraction on the first feature map output by the previous network, and sequentially undergo a 1×1 convolution, three 3×3 dilated convolutions, and a global average pooling operation, and then concatenate and pass a 1×1 convolution to obtain the second output feature map;
[0087] The first output feature map is concatenated with the second output feature map in the channel dimension as the input of the subsequent network.
[0088] In general, the present invention compresses the original U-Net model through structured pruning and asymmetric convolution technology, significantly reducing the number of parameters. At the same time, by introducing the coordinate attention module, the model's ability to capture target spatial features is improved. In addition, by cascading the improved U-Net network, a secondary learning opportunity is provided for the feature map to generate a more refined vascular probability map, further improving the segmentation performance. The lightweight cascade network model proposed in the present invention has only 0.70MB of parameters and achieves a high segmentation accuracy on the retinal image dataset.
[0089] In other embodiments, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the steps of the above-mentioned retinal image segmentation method based on a lightweight dual-path cascade network are implemented.
[0090] In other embodiments, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned retinal image segmentation method based on a lightweight two-way cascade network are implemented.
[0091] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0092] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A retinal image segmentation method based on a lightweight two-way cascade network, characterized in that: The following steps are involved: Two U-shaped networks with the same structure are introduced and improved as the front-end network and the back-end network respectively. The front-end network and the back-end network are cascaded to construct a retinal image segmentation model. In the previous network, the encoder part is performed in two stages. Each stage contains a residual module and a maximum pooling layer to downsample the original input retinal image. After one stage of processing, the resolution of the feature map is halved and the number of feature channels is doubled. In each stage, the feature map output by the residual module has two flow paths: one path continues to downsample to extract deeper semantic information; the other path enters the skip path, is processed by the coordinate attention module and connected to the decoder part; After the encoder part is completed, the feature map is processed by a residual module and then the decoding process begins; The decoding part consists of two stages. In each stage, upsampling is performed through deconvolution and residual modules to restore the size of the feature map and output the first feature map through 1×1 convolution. The first feature map output by the previous network is superimposed with the original input retinal image as the input of the subsequent network; The subsequent network processes the first feature map according to the same encoding and decoding process as the previous network to obtain the final retinal vascular image.
2. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 1 is characterized in that: The final retinal vascular image is represented as: U(x)=U( 2 )(x,U (1) (x)) Among them, U(x) represents the final retinal vascular image, x represents the original input retinal image, and U (1) (x) represents the first feature map output by the previous network, U (1) represents the previous network, U (2) Represents the subsequent network.
3. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 1 is characterized in that: The improvements to the U-type network include: using a structured pruning method to prune the five-layer architecture of the original U-type network to three layers, and pruning the feature channels of the original U-type network to {32, 64, 128}.
4. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 1 is characterized in that: The processing process of the coordinate attention module on the feature map of size W×H includes: Using pooling kernels of size (H, 1) and (1, W), the feature map of each channel is encoded along the x-axis and y-axis respectively to obtain the direction-aware feature map. The calculation formula is: in, represents the output of the cth channel with height h, and size is c×H×1, represents the output of the cth channel with a width of w, and the size is C×1×W; x(i,h) represents the value of the feature map with a width coordinate of i and a height of h along the cth channel in the x-axis direction, and x(w,j) represents the value of the feature map with a width of w and a height coordinate of j along the y-axis direction; Direction-aware feature maps via coordinate attention and Combined, the intermediate feature representation is generated through 1×1 convolution processing, and the calculation formula is: f∈δ(F1(Z h ·Z w )) Where · represents the connection operation in the spatial dimension, δ represents the nonlinear activation function, F1 represents the 1x1 convolution transformation function, and f∈R C / r*(H+w) Represents the intermediate feature map that encodes vertical and horizontal information; The intermediate feature representation f is decomposed into two directions: width and height, respectively: f h ∈R C / r*H and f w ∈R C / r*w , respectively, after 1×1 convolution F h 、F w And Sigmoid activation function δ is processed to match the number of channels of the input feature map to obtain g h and g w , the process is as follows: g h =δ(F h (f h )) g w =δ(F w (f w )) The calculated weight g h and g w Applied to the feature map x(i,j) after downsampling by the residual module and the maximum pooling layer, the final weighted feature representation Y=[y1,y2,…,y t ], so that the retinal image segmentation model focuses on the vascular area during training, where 5. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 1 is characterized in that: A dual-path information fusion module is introduced to cascade the previous network and the next network, and the first feature map output by the previous network is superimposed with the original retinal feature map.
6. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 5 is characterized in that: The step of superimposing the first feature map output by the previous network and the original retinal feature map by the dual-path information fusion module includes: The original input retinal image of the previous network is subjected to multi-scale feature extraction, and then subjected to one 1×1 convolution, three 3×3 dilated convolutions, and one global average pooling operation. After concatenation, the first output feature map is obtained through a 1×1 convolution. Perform multi-scale feature extraction on the first feature map output by the previous network, and sequentially undergo a 1×1 convolution, three 3×3 dilated convolutions, and a global average pooling operation, and then concatenate and pass a 1×1 convolution to obtain the second output feature map; The first output feature map is concatenated with the second output feature map in the channel dimension as the input of the subsequent network.
7. The retinal image segmentation method based on a lightweight dual-path cascade network according to claim 1 is characterized in that: Improvements to the U-type network also include: introducing asymmetric convolution.
8. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor; wherein when the processor executes the computer program, the steps of the retinal image segmentation method based on a lightweight dual-path cascade network as described in any one of claims 1 to 7 are implemented.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the retinal image segmentation method based on a lightweight two-way cascade network as described in any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Retinal vessel segmentation network and segmentation method based on improved Unet network, computer equipment and storage medium
CN115546227A
Remote sensing image building extraction method based on coordinate attention and data correlation upsampling
CN115631412A
Retinal vessel segmentation method based on mixed attention mechanism and asymmetric convolution
CN115731242A
Lightweight retinal vessel segmentation method based on attention mechanism
CN115760872A
Cited By
Road surface strip repair semantic segmentation method based on lightweight U-net
CN120510367A
