Image segmentation method fusing state modeling and convolution perception mechanism

Abstract

The invention discloses an image segmentation method fusing state modeling and a convolution perception mechanism. The method comprises the following steps: constructing a dual-path encoder through a convolution perception branch of a local perception path and a state space modeling branch of a global modeling path; inputting the acquired image data into a double-path encoder, and respectively extracting local semantic features and global semantic features of the image data by adopting a local sensing path and a global modeling path of the double-path encoder; constructing a space-semantic dynamic gating fusion module based on a gating unit, and performing feature fusion on the local semantic features and the global semantic features to obtain final fusion features; constructing an edge attention guidance decoder based on an edge attention guidance module, inputting the final fusion feature into the edge attention guidance decoder, reinforcing boundary information by using the edge attention guidance module, and finally outputting an image segmentation result by the edge attention guidance decoder; and high-precision segmentation of the image is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an image segmentation method that integrates state modeling and convolutional perception mechanisms. Background Technology

[0002] Skin lesion image segmentation is a particularly challenging task due to the diverse types of lesions, blurred boundaries, and complex color variations. Traditional Convolutional Neural Networks (CNNs), such as U-Net, improve segmentation through multi-scale feature fusion, but are limited by their receptive field and struggle to capture long-range dependencies. To enhance global modeling capabilities, Transformer structures have been introduced, utilizing self-attention mechanisms to capture long-range dependencies, but their high computational complexity limits their application to high-resolution images. State Space Models (SSMs) offer an efficient alternative with linear computational complexity and excellent long-range dependency modeling capabilities. Typical methods include VM-UNet, which enhances global semantic capture capabilities through a visual state space module; H-VMuNet introduces a high-order two-dimensional scanning module to enhance boundary recognition capabilities; and to balance global and local features, multi-branch structures have been proposed, with MRDB combining a visual state space model and ResNet to construct a dual encoder for extracting multi-dimensional features. Nevertheless, existing methods still have limitations in local detail perception, adaptive feature fusion, and boundary attention, making it difficult to meet the segmentation needs of complex boundaries in skin lesions. Most models focus on global semantic modeling, lacking sufficient ability to perceive lesion boundaries and detailed features, resulting in blurred edges and inaccurate detail representation in the segmentation results. Multi-branch feature fusion strategies are mostly static fusion or simple splicing, lacking dynamic adjustment mechanisms for the semantics of different regions, affecting the fusion effect and the adaptability of the model. Existing methods generally lack dedicated boundary enhancement modules, making it difficult to effectively handle lesion regions with blurred boundaries and complex shapes, affecting the accuracy of segmentation and clinical applicability. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides an image segmentation method that integrates state modeling and convolutional perception mechanism, constructs a dual-path encoder for perception modeling, and combines it with the strategy of spatial-semantic dynamic gating fusion module and edge attention-guided decoder structure to achieve high-precision image segmentation.

[0004] Technical Solution: To achieve the above objectives, the present invention provides an image segmentation method that integrates state modeling and convolutional perception mechanisms, comprising the following steps:

[0005] Step 1: Construct a dual-path encoder by using the convolutional perceptual branch of the local perceptual path and the state-space modeling branch of the global modeling path;

[0006] Step 2: Input the acquired image data into the dual-path encoder, and use the local perception path and global modeling path of the dual-path encoder to extract the local semantic features and global semantic features of the image data respectively;

[0007] Step 3: Construct a spatial-semantic dynamic gating fusion module based on gating units to fuse local semantic features and global semantic features to obtain the final fused features;

[0008] Step 4: Construct an edge attention guided decoder based on the edge attention guided module. Input the final fused features into the edge attention guided decoder and use the edge attention guided module to enhance the boundary information. Finally, the edge attention guided decoder outputs the image segmentation result.

[0009] Furthermore, in step one, the state-space modeling branch contains several state-space modeling layers; each state-space modeling layer in the state-space modeling branch introduces a state-space modeling module (SSM) to perform long-range modeling of the image data features; the state-space modeling module (SSM) uses a visual state-space module to capture long-range dependencies and global semantic information in the image; its mathematical form is as follows:

[0010]

[0011]

[0012] In the formula, Represents the state matrix, and All of these represent projection parameters.

[0013] Furthermore, each state space modeling layer in the state space modeling branch introduces a spatial awareness enhancement module, namely a shallow convolutional enhancement module, at the front end, and is deployed in series at the input position of the state space modeling module (SSM). The spatial awareness enhancement module includes an input layer, a convolutional layer, a two-dimensional normalization layer, a convolutional layer, a batch normalization layer, an activation function layer, and an output layer connected in sequence. Through continuous convolution and multiple normalization operations, the model's ability to represent texture edges and color boundaries of image data is enhanced. The overall calculation process of the spatial awareness enhancement module is shown in the following structure-aware function:

[0014]

[0015] In the formula, δ spatial Let θ represent the spatial structure enhancement operator, θ represent the learnable parameter set of the spatial perception enhancement module, and X represent the original input features.

[0016] Furthermore, each state space modeling layer in the state space modeling branch introduces a layer-by-layer convolutional refinement module at its end, and these modules are deployed in series at the output position of the state space modeling module (SSM). The layer-by-layer convolutional refinement module includes an input layer, a convolutional layer, a batch normalization layer, a nonlinear activation function layer, and an output layer connected in sequence. The layer-by-layer convolutional refinement module adopts a multi-level convolutional stacking structure to perform layer-by-layer local perception enhancement operations, thereby improving the reconstruction and discrimination capabilities of boundary details.

[0017] Furthermore, in step two, the acquired image data is input into a dual-path encoder, and the image data is input into both the local perception path and the global modeling path. The image data is input into the local perception path, where the convolutional perception branch includes four concatenated convolutional module layers. The image data is sequentially processed through these four convolutional module layers to extract four layers of local semantic features. The features extracted by the four convolutional module layers in the convolutional perception branch are e1 local semantic features, e2 local semantic features, e3 local semantic features, and e4 local semantic features.

[0018] Image data is input to the global modeling path. The state space modeling branch includes five cascaded state space modeling layers. Image data sequentially passes through these five layers for state space modeling operations. The input of each state space modeling layer is sequentially processed by the spatial perception enhancement module, the state space modeling module (SSM), and the layer-by-layer convolutional thinning module, ultimately extracting five layers of global semantic features. The features extracted by the five state space modeling layers in the state space modeling branch are v0 global semantic features, v1 global semantic features, v2 global semantic features, v3 global semantic features, and v4 global semantic features, respectively.

[0019] Furthermore, the spatial-semantic dynamic gating fusion module comprises four layers of spatial-semantic dynamic gating fusion modules; the e1 local semantic feature and the v1 global semantic feature are both input into the first layer of spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the first final fusion feature; the e2 local semantic feature and the v2 global semantic feature are both input into the second layer of spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the second final fusion feature; the e3 local semantic feature and the v3 global semantic feature are both input into the third layer of spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the third final fusion feature; and the e4 local semantic feature and the v4 global semantic feature are both input into the fourth layer of spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the fourth final fusion feature.

[0020] Furthermore, the spatial-semantic dynamic gating fusion module learns the semantic feature weights of different regions to achieve effective alignment and fusion of multi-scale features; it also integrates local semantic features F localand global semantic features F global The concatenated features, obtained along the channel dimension, are the fused input. The fused input is then processed sequentially through convolution and activation functions to generate a fused gating signal. The Softmax function is used to process the fused control signal to generate a fused gating map. Based on the fused gating map, the features at each position are weighted sequentially to output the final fused feature F. fusion The mathematical expression for the spatial-semantic dynamic gating fusion module is as follows:

[0021]

[0022]

[0023] In the formula, C g This indicates a gated signal generation convolutional network, C f This indicates that the feature mapping adjusts the network. represents pixel-wise multiplication; G represents the fusion gating map, controlling the weight of the fused features at each location; F fusion This indicates the final fusion feature.

[0024] Furthermore, in step four, the edge attention-guided decoder includes four layers of decoders: D1 decoder, D2 decoder, D3 decoder, and D4 decoder; both D4 decoder and D3 decoder incorporate edge attention-guided modules; the output of D4 decoder is input to D3 decoder for decoding, the output of D3 decoder is input to D2 decoder for decoding, the output of D2 decoder is input to D1 decoder for decoding, and the image segmentation result is output after D1 decoder finishes decoding.

[0025] Furthermore, in the edge attention guidance module, local response features are first extracted, and then processed using a sigmoid activation function to generate an edge attention map. Based on the edge attention map, gated multiplication is used to control the intensity of the output feature response, finally obtaining the enhanced features of the edge region response. Then, the enhanced features of the edge region response are transmitted to the downstream decoder, and the input feature of the edge attention guidance module is set to F. in Its mathematical expression is as follows:

[0026]

[0027]

[0028] In the formula, Conv() represents the edge extraction convolution operation; σ represents the Sigmoid activation function, which compresses the response to [0, 1]; E represents the edge attention map. Indicates pixel-wise multiplication; F out This represents the enhanced response characteristics of the edge region.

[0029] Beneficial Effects: This invention provides an image segmentation method that integrates state modeling and convolutional perception mechanisms. It constructs a dual-path encoder for perception modeling, combined with a spatial-semantic dynamic gating fusion module and an edge attention-guided decoder structure, achieving high-precision image segmentation. The dual-path encoder simultaneously extracts local details and global semantics, enhancing feature representation capabilities. A spatial perception enhancement module is introduced to strengthen local structure perception in the state-space modeling branch. A spatial-semantic dynamic gating fusion module is designed to achieve adaptive fusion of multi-path features, improving fusion performance. The edge attention-guided module enhances the response to lesion boundaries, improving segmentation accuracy and boundary restoration capabilities. The state-space modeling module efficiently models long-distance dependencies, balancing performance and computational efficiency. This significantly overcomes the shortcomings of existing methods in capturing local details and dynamic fusion, improving the accuracy and practicality of skin lesion segmentation. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the network structure of the image segmentation method of the present invention;

[0031] Figure 2 A schematic diagram of the structure of any state space modeling layer in the state space modeling branch;

[0032] Figure 3 This is a schematic diagram of the spatial perception enhancement module.

[0033] Figure 4 This is a schematic diagram illustrating the dynamic fusion of local and global semantic features.

[0034] Figure 5 This is a structural diagram of the spatial-semantic dynamic gating fusion module;

[0035] Figure 6 This is a schematic diagram of the edge attention guidance module.

[0036] Figure 7 Visualization of image segmentation results using this invention on the ISIC 2017 and PH2 datasets;

[0037] Figure 8 A visual comparison of the segmentation results of this invention and other image segmentation methods;

[0038] Figure 9 A bar chart comparing the segmentation results of the ISIC2017 dataset;

[0039] Figure 10 A bar chart comparing the segmentation results of the PH2 dataset. Detailed Implementation

[0040] The invention will now be further described with reference to the accompanying drawings.

[0041] like Figure 1 As shown, an image segmentation method that integrates state modeling and convolutional perception mechanisms includes the following steps:

[0042] Step 1: Construct a dual-path encoder by using the convolutional perceptual branch of the local perceptual path and the state-space modeling branch of the global modeling path;

[0043] Step 2: Input the acquired image data into the dual-path encoder. Use the local perception path and global modeling path of the dual-path encoder to extract the local semantic features and global semantic features of the image data, respectively. This is to obtain the local path features and global path features of the image data.

[0044] Step 3: Construct a spatial-semantic dynamic gating fusion module based on gating units to fuse local semantic features and global semantic features to obtain the final fused features;

[0045] Step 4: Construct an edge attention guided decoder based on the edge attention guided module. Input the final fused features into the edge attention guided decoder and use the edge attention guided module to enhance the boundary information. Finally, the edge attention guided decoder outputs the image segmentation result.

[0046] like Figure 2 As shown, in step one, the state space modeling branch contains several state space modeling layers; each state space modeling layer in the state space modeling branch introduces a state space modeling module (SSM) to perform long-distance modeling of the image data features; the state space modeling module (SSM) uses a visual state space module to capture long-distance dependencies and global semantic information in the image; its mathematical form is as follows:

[0047]

[0048]

[0049] In the formula, Represents the state matrix, and Both represent projection parameters; h(t) represents the hidden state of the system. Let y(t) represent the time derivative of the hidden state, x(t) represent the features of the input image, and y(t) represent the output result.

[0050] like Figure 1-3As shown, to overcome the problem of insufficient perception of local details in the State Space Modeling Module (SSM), a spatial awareness enhancement module, namely a shallow convolutional enhancement module, is introduced at the front end of each state space modeling layer in the state space modeling branch, and is deployed in series at the input position of the SSM. The spatial awareness enhancement module includes an input layer, a convolutional layer, a two-dimensional normalization layer, a convolutional layer, a batch normalization layer, an activation function layer, and an output layer connected in sequence. Through continuous convolution and multiple normalization operations, the model's ability to represent key regions such as texture edges and color boundaries of image data is enhanced without significantly increasing the computational load. The spatial awareness enhancement module as a whole can be regarded as a lightweight local structure extractor, mainly composed of two convolutional layers, two-dimensional normalization, and batch normalization, and nonlinear modeling capability is introduced through the activation function. Its function is equivalent to performing a pre-encoding process on the original input feature X, improving the SSM's response to spatial structure. The overall calculation process of the spatial awareness enhancement module is shown in the following structure-aware function:

[0051]

[0052] In the formula, δ spatial Let θ represent the spatial structure enhancement operator, θ represent the set of learnable parameters of the spatial perception enhancement module, and X represent the original input features, or the original input image data.

[0053] like Figure 2 As shown, to address the issues of ambiguous output feature representation and weak structural edge response in the State Space Modeling (SSM) module, a layer-by-layer convolutional refinement module is introduced at the end of each state space modeling layer in the state space modeling branch, and deployed in series at the output position of the SSM module. The layer-by-layer convolutional refinement module has a simple structure, mainly including continuous convolution, batch normalization, and nonlinear activation operations. It comprises an input layer, a convolutional layer, a batch normalization layer, a nonlinear activation function layer, and an output layer connected in sequence. The module employs a multi-level convolutional stacking structure to perform layer-by-layer local perception enhancement operations, thereby improving the reconstruction and discrimination capabilities of boundary details. The layer-by-layer convolutional refinement module possesses the functions of refining, compressing, and nonlinearly mapping boundary features. Its function is similar to an image edge enhancer, compensating for the structural shortcomings of the SSM module's insufficient response to spatial boundaries.

[0054] like Figure 1As shown, in step two, the acquired image data is input into a dual-path encoder, and the image data is input into both the local perception path and the global modeling path. The image data is input into the local perception path, where the convolutional perception branch includes four concatenated convolutional module layers. The image data sequentially passes through these four convolutional module layers for convolutional operations, extracting four layers of local semantic features. The features extracted by the four convolutional module layers in the convolutional perception branch are e1 local semantic features, e2 local semantic features, e3 local semantic features, and e4 local semantic features. The convolutional perception branch uses a multi-layer convolutional module based on a ResNet residual network, responsible for extracting local texture and detail information from the input image. This method uses four convolutional module layers. Image data is input into the first convolutional module layer CNN1, which extracts features through convolution to obtain local semantic features e1. The local semantic features e1 are then input into the second convolutional module layer CNN2, which extracts features through convolution to obtain local semantic features e2. The local semantic features e2 are then input into the third convolutional module layer CNN3, which extracts features through convolution to obtain local semantic features e3. The local semantic features e3 are then input into the fourth convolutional module layer CNN4, which extracts features through convolution to obtain local semantic features e4.

[0055] like Figure 1As shown, image data is input to the global modeling path. The state space modeling branch includes five cascaded state space modeling layers. Image data sequentially passes through these five layers for state space modeling operations. The input of each state space modeling layer sequentially passes through the spatial perception enhancement module, the state space modeling module (SSM), and the layer-by-layer convolutional thinning module, ultimately extracting five layers of global semantic features. The features extracted by the five state space modeling layers in the state space modeling branch are v0 global semantic features, v1 global semantic features, v2 global semantic features, v3 global semantic features, and v4 global semantic features, respectively. Image data is input into the first state-space modeling layer VSS0, where state-space modeling is performed to obtain the global semantic feature v0. The global semantic feature v0 is then input into the second state-space modeling layer VSS1, where state-space modeling is performed to obtain the global semantic feature v1. The global semantic feature v1 is then input into the third state-space modeling layer VSS2, where state-space modeling is performed to obtain the global semantic feature v2. The global semantic feature v2 is then input into the fourth state-space modeling layer VSS3, where state-space modeling is performed to obtain the global semantic feature v3. The global semantic feature v3 is then input into the fifth state-space modeling layer VSS4, where state-space modeling is performed to obtain the global semantic feature v4.

[0056] The specific process of each state-space modeling layer is as follows: Image features are input to the spatial perception enhancement module. Through continuous convolution and multiple normalization operations, the model's ability to represent key regions such as texture edges and color boundaries in the image data is strengthened. Then, the processed image data is input to the state-space modeling module (SSM). The SSM performs long-distance modeling on the image features, and then the processed image data is input to the layer-by-layer convolutional refinement module. The layer-by-layer convolutional refinement module adopts a multi-level convolution stacked structure to perform layer-by-layer local perception enhancement operations, ultimately obtaining global semantic features. Then, both local and global semantic features are input to the spatial-semantic dynamic gating fusion module for adaptive weighted fusion at the spatial and semantic levels to obtain the final fused features.

[0057] like Figure 3 As shown, the image features input to the spatial awareness enhancement module undergo a series of operations: input layer, convolutional layer, 2D normalization layer, convolutional layer, batch normalization layer, activation function layer, and output layer. The processed image is then input to the state space modeling module (SSM). Similarly, the image features input to the layer-by-layer convolutional thinning module undergo a series of operations: input layer, convolutional layer, batch normalization layer, nonlinear activation function layer, and output layer, ultimately outputting global semantic features.

[0058] like Figure 4 As shown, the spatial-semantic dynamic gating fusion module comprises four layers of spatial-semantic dynamic gating fusion modules. The e1 local semantic feature and the v1 global semantic feature are both input into the first layer of the spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the first final fusion feature. The e2 local semantic feature and the v2 global semantic feature are both input into the second layer of the spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the second final fusion feature. The e3 local semantic feature and the v3 global semantic feature are both input into the third layer of the spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the third final fusion feature. The e4 local semantic feature and the v4 global semantic feature are both input into the fourth layer of the spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the fourth final fusion feature. The local and global semantic features of each layer are input into the spatial-semantic dynamic gating fusion module. First, they are concatenated. Then, the concatenated features are sequentially processed through convolutional layer operations, activation function operations, convolutional layer operations, and Softmax function operations to output a fusion gating map. Finally, the features at each position are weighted according to the fusion gating map to obtain the final fusion feature of that layer.

[0059] like Figure 5 As shown, the spatial-semantic dynamic gating fusion module learns the semantic feature weights of different regions to achieve effective alignment and fusion of multi-scale features; it also achieves semantic alignment and fusion control of feature maps of different paths, improving the overall feature representation capability; the spatial-semantic dynamic gating fusion module includes a gating unit, used to perform adaptive weighted fusion of local and global semantic features output by the dual-path encoder at the spatial and semantic levels. The spatial-semantic dynamic gating fusion module will convert the local semantic features F local and global semantic features F global The concatenated features, obtained along the channel dimension, are the fused input. The fused input is then processed sequentially through convolution and activation functions to generate a fused gating signal. The Softmax function is used to process the fused control signal to generate a fused gating map. Based on the fused gating map, the features at each position are weighted sequentially to output the final fused feature F. fusion The mathematical expression for the spatial-semantic dynamic gating fusion module is as follows:

[0060]

[0061]

[0062] In the formula, C g This indicates a gated signal generation convolutional network, C f This indicates that the feature mapping adjusts the network. represents pixel-wise multiplication; G represents the fusion gating map, controlling the weight of the fused features at each location; F fusion This indicates the final fusion feature.

[0063] like Figure 1 As shown, in step four, the edge attention-guided decoder includes four layers of decoders: D1 decoder, D2 decoder, D3 decoder, and D4 decoder. Both D4 decoder and D3 decoder incorporate edge attention-guided modules. The output of D4 decoder is input to D3 decoder for decoding, the output of D3 decoder is input to D2 decoder for decoding, the output of D2 decoder is input to D1 decoder for decoding, and the image segmentation result is output after D1 decoder finishes decoding.

[0064] The fourth final fusion feature is input to the D4 decoder as its input feature. The D4 decoder decodes the input feature to obtain its output. The output of the D4 decoder is input to the D3 decoder, and the third final fusion feature is fused with the output of the D4 decoder through a residual skip connection to obtain the input feature of the D3 decoder. The D3 decoder decodes the input feature to obtain its output. The output of the D3 decoder is input to the D2 decoder, and the second final fusion feature is fused with the output of the D3 decoder through a residual skip connection to obtain the input feature of the D2 decoder. The D2 decoder decodes the input feature to obtain its output. The output of the D2 decoder is input to the D1 decoder, and the first final fusion feature is fused with the output of the D2 decoder through a residual skip connection to obtain the input feature of the D1 decoder. The D1 decoder decodes the input feature to obtain its output, which is the image segmentation result. The residual skip connection enhances the detail restoration capability of the image features and effectively fuses the low-level features of the encoder into the decoding path, achieving the collaborative recovery of structural and semantic information.

[0065] Furthermore, edge attention guidance modules are introduced in both the D4 and D3 decoders. Therefore, before decoding the input features, the edge attention guidance module first processes the input features to obtain features with enhanced edge region responses, and then decodes these enhanced features to obtain the output of the D4 decoder. Similarly, before decoding the input features, the edge attention guidance module first processes the input features to obtain features with enhanced edge region responses, and then decodes these enhanced features to obtain the output of the D3 decoder. The edge attention guidance module employs an edge attention mechanism. Structurally, the edge attention guidance module and each decoder layer are independent of each other, but functionally connected sequentially, forming the edge enhancement decoding stage.

[0066] like Figure 6 As shown, the edge attention guidance module focuses on responding to the lesion boundary region in the input features; it uses edge information to guide the decoding process, improving the restoration quality and structural integrity of blurred boundaries; it strengthens the image boundary response through the edge attention mechanism, improving segmentation accuracy; then the decoder gradually restores the spatial resolution, generating the final high-precision lesion segmentation map, which is the image segmentation result. In the edge attention guidance module, local response features are first extracted, then processed using the Sigmoid activation function to generate an edge attention map; based on the edge attention map, gated multiplication is used to control the intensity of the output feature response, finally obtaining the enhanced edge region response features; then the enhanced edge region response features are transmitted to the downstream decoder, and the input feature of the edge attention guidance module is set as F. in Its mathematical expression is as follows:

[0067]

[0068]

[0069] In the formula, Conv() represents the edge extraction convolution operation; σ represents the Sigmoid activation function, which compresses the response to [0, 1]; E represents the edge attention map. This represents pixel-by-pixel multiplication, enabling spatial guidance; F out This represents the enhanced response characteristics of the edge region; .

[0070] like Figure 6 As shown, the edge attention guidance module includes an input layer, a convolutional layer, a sigmoid activation function layer, and an output layer. Image features are input to the input layer, then to the convolutional layer for convolution. The convolutional layer then inputs the convolutional features to the sigmoid activation function layer, which processes the features to obtain an output. The output of the sigmoid activation function layer is then multiplied with the convolutional features by a gated multiplication operation. The result of the gated multiplication is input to the output layer, which outputs the enhanced edge region features to the downstream decoder path for further decoding, ultimately yielding the image segmentation result.

[0071] By employing a dual-path encoder and a spatial-semantic dynamic gating fusion module, both local texture and global contextual information are considered, significantly improving the ability to express complex lesions. The edge guidance module effectively enhances boundary sensitivity, improving the accuracy and structural integrity of the segmentation results. Therefore, the image segmentation method of this invention is compared with the U-Net, VM-UNet, MRDB, and H-vmunet algorithms on the ISIC2017 and PH2 datasets for image segmentation. The segmentation results of the image segmentation method of this invention are compared with those of the U-Net, VM-UNet, MRDB, and H-vmunet algorithms. Figure 7-8 As shown; simultaneously, a bar chart comparing the Dice coefficient and the IoU index is created, as shown. Figure 9-10 As shown.

[0072] like Figure 7 As shown, Figure 7 (a) is the original image. Figure 7 (b) is the true label. Figure 7 (c) is the image segmentation result of the present invention; such as Figure 8 As shown, Figure 8 (a) is the original image. Figure 8 (b) is the true label. Figure 8 (c) shows the image segmentation result of the present invention. Figure 8 (d) shows the segmentation result of the U-Net algorithm. Figure 8 (e) shows the segmentation result of the VM-UNet algorithm. Figure 8 (f) shows the segmentation result of the MRDB algorithm. Figure 8 (g) shows the segmentation result of the H-vmunet algorithm; the comparison results show that the image segmentation effect of the present invention is better.

[0073] like Figure 9 As shown, after image segmentation of the ISIC2017 dataset, the Dice coefficient of the image segmentation result in this scheme is 0.9225, which is higher than that of other segmentation algorithms. Furthermore, the false negative rate of the image segmentation result in this scheme is 0.0573, which is lower than that of other segmentation algorithms. Figure 10 As shown, after segmenting the images in the PH2 dataset, the Dice coefficient of the image segmentation result in this scheme is 0.9203, which is higher than that of other segmentation algorithms. Furthermore, the false negative rate of the image segmentation result in this scheme is 0.0923, which is the lowest compared to other segmentation algorithms. This demonstrates that the accuracy of the image segmentation result in this invention is higher than that of other segmentation algorithms.

[0074] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. An image segmentation method that integrates state modeling and convolutional perception mechanisms, characterized in that: Includes the following steps: Step 1: Construct a dual-path encoder by using the convolutional perceptual branch of the local perceptual path and the state-space modeling branch of the global modeling path; Step 2: Input the acquired image data into the dual-path encoder, and use the local perception path and global modeling path of the dual-path encoder to extract the local semantic features and global semantic features of the image data respectively; Step 3: Construct a spatial-semantic dynamic gating fusion module based on gating units to fuse local semantic features and global semantic features to obtain the final fused features; Step 4: Construct an edge attention guided decoder based on the edge attention guided module. Input the final fused features into the edge attention guided decoder and use the edge attention guided module to enhance the boundary information. Finally, the edge attention guided decoder outputs the image segmentation result.

2. The image segmentation method according to claim 1, which integrates state modeling and convolutional perception mechanisms, is characterized in that: In step one, the state-space modeling branch contains several state-space modeling layers; each state-space modeling layer in the state-space modeling branch introduces a state-space modeling module (SSM) to perform long-range modeling of the features of the image data; the state-space modeling module (SSM) uses a visual state-space module to capture long-range dependencies and global semantic information in the image; its mathematical form is as follows: In the formula, Represents the state matrix, and All of these represent projection parameters.

3. The image segmentation method according to claim 2, which integrates state modeling and convolutional perception mechanisms, is characterized in that: Each state-space modeling layer in the state-space modeling branch introduces a spatial awareness enhancement module (SSM), also known as a shallow convolutional enhancement module, at the front end of the SSM. This SSM is deployed in series at the input position of the SSM. The SSM comprises an input layer, a convolutional layer, a 2D normalization layer, another convolutional layer, a batch normalization layer, an activation function layer, and an output layer, connected sequentially. Through continuous convolution and multiple normalization operations, the model's ability to represent texture edges and color boundaries in image data is enhanced. The overall computation process of the SSM is shown in the following structure-aware function: In the formula, δ spatial Let θ represent the spatial structure enhancement operator, θ represent the learnable parameter set of the spatial perception enhancement module, and X represent the original input features.

4. The image segmentation method according to claim 2, which integrates state modeling and convolutional perception mechanisms, is characterized in that: Each state space modeling layer in the state space modeling branch introduces a layer-by-layer convolutional refinement module at its end, and these modules are deployed in series at the output position of the state space modeling module (SSM). The layer-by-layer convolutional refinement module includes an input layer, a convolutional layer, a batch normalization layer, a nonlinear activation function layer, and an output layer connected in sequence. The layer-by-layer convolutional refinement module adopts a multi-level convolutional stacking structure to perform layer-by-layer local perception enhancement operations, thereby improving the reconstruction and discrimination capabilities of boundary details.

5. The image segmentation method according to claim 1, which integrates state modeling and convolutional perception mechanisms, is characterized in that: In step two, the acquired image data is input into a dual-path encoder, and the image data is input into both the local perception path and the global modeling path. The image data is input into the local perception path, where the convolutional perception branch includes four concatenated convolutional module layers. The image data is sequentially processed through these four convolutional module layers to extract four layers of local semantic features. The features extracted by the four convolutional module layers in the convolutional perception branch are e1 local semantic features, e2 local semantic features, e3 local semantic features, and e4 local semantic features. Image data is input to the global modeling path. The state space modeling branch includes five consecutive state space modeling layers. Image data passes through the five state space modeling layers in sequence to perform state space modeling operations. The input of each state space modeling layer passes through the spatial perception enhancement module, the state space modeling module SSM, and the layer-by-layer convolutional refinement module in sequence to finally extract the five layers of global semantic features. The features extracted from the five state space modeling layers in the state space modeling branch are v0 global semantic features, v1 global semantic features, v2 global semantic features, v3 global semantic features, and v4 global semantic features.

6. The image segmentation method according to claim 5, which integrates state modeling and convolutional perception mechanisms, is characterized in that: The spatial-semantic dynamic gating fusion module includes four layers of spatial-semantic dynamic gating fusion modules; the e1 local semantic features and v1 global semantic features are both input into the first layer of spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the first final fusion feature; The e2 local semantic features and v2 global semantic features are both input into the second-layer spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the second final fused feature; the e3 local semantic features and v3 global semantic features are both input into the third-layer spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the third final fused feature; the e4 local semantic features and v4 global semantic features are both input into the fourth-layer spatial-semantic dynamic gating fusion module for dynamic adaptive fusion to obtain the fourth final fused feature.

7. The image segmentation method according to claim 6, which integrates state modeling and convolutional perception mechanisms, is characterized in that: The spatial-semantic dynamic gating fusion module learns the semantic feature weights of different regions to achieve effective alignment and fusion of multi-scale features; it also integrates local semantic features F local and global semantic features F global The concatenated features, obtained along the channel dimension, are the fused input. The fused input is then processed sequentially through convolution and activation functions to generate a fused gating signal. The Softmax function is used to process the fused control signal to generate a fused gating map. Based on the fused gating map, the features at each position are weighted sequentially to output the final fused feature F. fusion The mathematical expression for the spatial-semantic dynamic gating fusion module is as follows: In the formula, C g This indicates a gated signal generation convolutional network, C f This indicates that the feature mapping adjusts the network. represents pixel-wise multiplication; G represents the fusion gating map, which controls the weight of the fused features at each location; F fusion This indicates the final fusion feature.

8. The image segmentation method according to claim 1, which integrates state modeling and convolutional perception mechanisms, is characterized in that: In step four, the edge attention-guided decoder includes four layers of decoders: D1 decoder, D2 decoder, D3 decoder, and D4 decoder; both the D4 decoder and the D3 decoder incorporate an edge attention-guided module. The output of the D4 decoder is input to the D3 decoder for decoding, the output of the D3 decoder is input to the D2 decoder for decoding, the output of the D2 decoder is input to the D1 decoder for decoding, and the D1 decoder outputs the image segmentation result after decoding.

9. The image segmentation method according to claim 8, which integrates state modeling and convolutional perception mechanisms, is characterized in that: In the edge attention guidance module, local response features are first extracted, then processed using a sigmoid activation function to generate an edge attention map. Based on the edge attention map, gated multiplication is used to control the intensity of the output feature response, finally obtaining the enhanced features of the edge region response. These enhanced features are then transmitted to the downstream decoder, with the input feature of the edge attention guidance module set to F. in Its mathematical expression is as follows: In the formula, Conv() represents the edge extraction convolution operation; σ represents the Sigmoid activation function, which compresses the response to [0, 1]; E represents the edge attention map. Indicates pixel-wise multiplication; F out This represents the enhanced response characteristics of the edge region.

Images