Target object segmentation method and system

Abstract

The invention provides a target object segmentation method and system, and relates to the technical field of image processing, and the method comprises the steps: obtaining a detection image containing a target object, and carrying out the feature extraction of the detection image, and obtaining a bottom feature and an up-sampling feature; fusing the underlying features and the up-sampling features through a cascade aggregation module to obtain a cascade feature map; converting the underlying features into a boundary feature map through a boundary extraction module, and respectively performing boundary extension and semantic extension on the boundary feature map and the cascade feature map to obtain a final boundary feature map and a final cascade feature map; and predicting the target object according to the final boundary feature map and the final cascade feature map to obtain a segmentation result of the target object. According to the method, the boundary details and the overall semantic information of the target object are comprehensively considered, so that the segmentation accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a target object segmentation method and system. Background Art

[0002] In the field of image processing technology, it is difficult to segment target objects, especially those with unclear contours. For example, in the medical field, the segmentation of polyp images is more difficult because the boundaries of polyps are not obvious and the contrast is low. Therefore, the attention mechanism is usually used to segment the contours of target objects.

[0003] In the prior art, the contour of the target object is segmented through the attention mechanism, which usually has a good effect on the segmentation of objects with clear edges, such as middle and late stage polyps with high contrast, but has a poor effect on objects with low edge contrast, such as early stage polyps. For Polyp-PVT, it is usually difficult to effectively fuse high-level semantic information with shallow boundary information due to the high fuzziness of the boundaries of some objects, such as small-sized polyps, resulting in poor effect on the segmentation of object contours. Summary of the invention

[0004] The problem solved by the present invention is how to improve the accuracy of target object contour segmentation.

[0005] In order to solve the above problems, the present invention provides a target object segmentation method and system.

[0006] In a first aspect, the present invention provides a target object segmentation method, comprising:

[0007] Acquire a detection image containing a target object, and perform feature extraction on the detection image to obtain underlying features and up-sampled features;

[0008] The bottom-level features and the up-sampled features are fused through a cascade aggregation module to obtain a cascade feature map;

[0009] The bottom layer features are converted into a boundary feature map by a boundary extraction module, and the boundary feature map and the cascade feature map are respectively subjected to boundary expansion and semantic expansion to obtain a boundary final feature map and a cascade final feature map;

[0010] The target object is predicted according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

[0011] Optionally, extracting features from the detection image to obtain underlying features and up-sampled features includes:

[0012] Performing deep feature extraction on the detection image through an encoder network of a pyramid vision transformer to obtain a bottom-layer feature map and a deep-layer feature map;

[0013] The deep feature map is upsampled by an upsampling module of the encoder network to obtain an upsampled feature map with the same resolution as the original detection image;

[0014] The underlying feature map is combined with the upsampled feature map to obtain the underlying features and the corresponding upsampled features.

[0015] Optionally, fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map includes:

[0016] The underlying features and the upsampled features are fused through the lightweight attention mechanism of the cascade aggregation module to obtain an attention-weighted feature map;

[0017] The underlying features are weighted from the channel dimension and the spatial dimension through the channel space attention module of the cascade aggregation module to obtain a channel attention feature map and a spatial attention feature map;

[0018] The cascade feature map is obtained according to the attention weighted feature map, the channel attention feature map and the spatial attention feature map.

[0019] Optionally, the fusing of the underlying features and the upsampling through the lightweight attention mechanism of the cascade aggregation module to obtain an attention-weighted feature map includes:

[0020] Performing a convolution operation on the upsampled features and the underlying features of the detection image to obtain convolution operation results corresponding to the upsampled features and the underlying features respectively;

[0021] Performing normalization processing according to the convolution operation results corresponding to the up-sampled features and the underlying features respectively;

[0022] The normalized convolution operation result is subjected to light attention allocation through the ReLU layer activation function and the Sigmoid activation function in turn to obtain an attention-weighted feature map.

[0023] Optionally, the channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including:

[0024] Through the channel space attention module, the underlying features are pooled using an adaptive maximum pooling function and an adaptive average pooling function;

[0025] The channel attention feature map is determined based on the underlying features after the pooling operation through the Sigmoid activation function and the Hadamard product.

[0026] Optionally, the channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including:

[0027] Through the channel space attention module, channel connection is performed according to the maximum value and average value obtained along the channel dimension of the underlying features;

[0028] The spatial attention feature map is determined by the Sigmoid activation function and the Hadamard product according to the underlying features after channel connection.

[0029] Optionally, the converting the underlying features into a boundary feature map by a boundary extraction module includes:

[0030] Inputting the bottom-level features into a preset convolutional layer of the boundary extraction module, and obtaining two adjacent feature cascades by reducing the channel size;

[0031] Inputting the features into a two-stream network in cascade order to obtain a plurality of boundary features, wherein each stream in the two-stream network uses a different convolution kernel;

[0032] The boundary features are combined by element-wise multiplication to obtain the boundary feature map.

[0033] Optionally, performing boundary expansion and semantic expansion on the boundary feature map and the cascade feature map respectively to obtain a boundary final feature map and a cascade final feature map includes:

[0034] Separating the boundary region and the internal region of the target object from the cascade feature map by using a morphological operator;

[0035] Fusing the boundary region and the internal region of the object to obtain a final feature map of the boundary;

[0036] Extracting high-level semantic features from the cascaded feature maps;

[0037] The cascade feature map is expanded according to the high-level semantic features through a cross non-local network to obtain the cascade final feature map.

[0038] Optionally, predicting the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object includes:

[0039] Inputting the boundary final feature map and the cascade final feature map into a preset convolutional layer, predicting the target object through the preset convolutional layer, and obtaining prediction results corresponding to the boundary final feature map and the cascade final feature map respectively;

[0040] The prediction results corresponding to the boundary final feature map and the cascade final feature map are summed to obtain the segmentation result of the target object.

[0041] In a second aspect, the present invention further provides a target object segmentation system, comprising:

[0042] An acquisition unit is used to acquire a detection image containing a target object, and perform feature extraction on the detection image to obtain underlying features and up-sampled features;

[0043] A fusion unit, used for fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map;

[0044] An expansion unit, configured to convert the underlying features into a boundary feature map through a boundary extraction module and perform boundary expansion and semantic expansion on the boundary feature map and the cascade feature map, respectively, to obtain a boundary final feature map and a cascade final feature map;

[0045] A prediction unit is used to predict the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

[0046] The target object segmentation method and system of the present invention obtains a detection image containing a target object and extracts underlying features and upsampled features from it, and these features include the shape, texture and context information of the target object. Then, the underlying features are fused with the upsampled features by using a cascade aggregation module to generate a cascade feature map containing rich details and semantic information. At the same time, the boundary extraction module converts the underlying features into a boundary feature map, focusing on the edge information of the target object. Then, by performing boundary extension and semantic extension on the boundary feature map and the cascade feature map respectively, the boundary information and high-level semantic information in the feature map are further strengthened, and the quality of the feature map is further improved, so that the boundary final feature map and the cascade final feature map can more accurately represent the contour and internal structure of the target object. Finally, the target object is predicted in combination with the boundary final feature map and the cascade final feature map to obtain a segmentation result. The present invention effectively improves the accuracy of segmentation by comprehensively considering the boundary details and overall semantic information of the target object, especially when processing a target object with unclear boundaries or low contrast, it can better identify and segment the contour of the target object, such as segmenting polyps with unclear boundaries and low contrast in the medical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a target object segmentation method according to an embodiment of the present invention;

[0048] Figure 2 FIG2 is a second flow chart of the target object segmentation method according to an embodiment of the present invention;

[0049] Figure 3 FIG. 4 is a structural block diagram of a target object segmentation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0051] Combination Figure 1 As shown, the present invention provides a target object segmentation method, comprising:

[0052] A detection image containing a target object is obtained, and features are extracted from the detection image to obtain underlying features and up-sampled features.

[0053] Specifically, detection images containing target objects are obtained. These images may come from different scenes and conditions and contain various types of target objects. The quality, resolution and characteristics of the image and the target object may vary. In a preferred embodiment of the present invention, in the field of medical image segmentation, such as polyp segmentation, this step involves identifying the location of polyps from images obtained by endoscopy. These images may contain multiple types of polyps with different shapes, sizes and contrasts, providing necessary input data for subsequent feature extraction and segmentation tasks. After obtaining the detection image, the underlying features and upsampled features are obtained through a feature extraction process. The underlying features usually contain more details and edge information, while the upsampled features contain higher-level semantic information.

[0054] The underlying features and the up-sampled features are fused through a cascade aggregation module to obtain a cascade feature map.

[0055] Specifically, the role of the cascade aggregation module is to effectively fuse the underlying features and upsampled features to generate a cascade feature map containing rich details and semantic information. In this step, through mechanisms such as lightweight attention and channel space attention modules, information transmission between different feature levels is promoted to achieve the extraction of high-level semantic information, thereby obtaining a feature map that integrates multi-scale information. It is worth mentioning that before the cascade aggregation module is used, it is necessary to first construct a cascade aggregation module. In a preferred embodiment of the present invention, the cascade aggregation module consists of two cascade parts, including lightweight attention and channel space attention.

[0056] The underlying features are converted into a boundary feature map through a boundary extraction module, and boundary expansion and semantic expansion are performed on the boundary feature map and the cascade feature map respectively to obtain a boundary final feature map and a cascade final feature map.

[0057] Specifically, the purpose of the boundary extraction module is to extract the boundary information of the target object from the underlying features and generate a boundary feature map. In this step, a camouflage removal block is introduced to remove the camouflage around the polyps, and the boundary features across levels and multiple scales are captured through a two-stream network and different convolution kernels, thereby obtaining a feature map that highlights the boundary of the target object. After obtaining the boundary feature map and the cascade feature map, the present invention further enhances the information of these two feature maps through boundary expansion and semantic expansion. Boundary expansion separates the boundary and internal area through morphological operators, while semantic expansion uses cross-non-local networks and Transformer attention mechanisms to strengthen high-level semantic features, so that the boundary final feature map and the cascade final feature map can more accurately represent the boundary and internal structure of the target object.

[0058] The target object is predicted according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

[0059] Specifically, the target object is predicted based on the boundary final feature map and the cascade final feature map to obtain the segmentation result. In this step, the feature map that combines boundary and semantic information provides rich context and location information, so that the network can more accurately identify and segment the target object, especially when dealing with target objects with unclear boundaries and low contrast, such as polyps in medical images, thereby improving the accuracy and robustness of segmentation. Through this method of comprehensively considering boundary and semantic information, the present invention can effectively improve the quality of target object segmentation.

[0060] In a preferred embodiment of the present invention, Figure 2 As shown in the figure, the image segmentation process starts with the image detection unit obtaining a detection image containing the target object. Then the feature extraction unit performs deep feature extraction on the image through the pyramid visual transformer encoder network to generate the underlying feature map. The convolution layer and the pooling layer work together to reduce the spatial dimension of the image while increasing the feature depth and semantic information. The upsampling module upsamples the deep feature map to the same resolution as the original image to generate an upsampled feature map. These feature maps are sent to the cascade aggregation module, where the lightweight attention mechanism and the channel space attention module process the feature maps respectively to enhance the feature representation. The lightweight attention mechanism assigns attention weights through convolution operations, normalization, and ReLU and Sigmoid activation functions, while the channel space attention module determines the channel attention feature map through adaptive pooling and Hadamard product. Cascade feature map and boundary feature Figure 1 Then, the morphological operator of the morphological fusion module separates the boundary area and the internal area of ​​the object from the cascade feature map, and semantically expands the cascade feature map through the cross non-local network. At the same time, the boundary area and the internal area are fused to obtain the boundary final feature map. Finally, the target object is predicted based on the boundary final feature map and the cascade final feature map to obtain the segmentation result.

[0061] The target object segmentation method of the present invention obtains a detection image containing a target object and extracts underlying features and upsampled features from it, and these features include the shape, texture and context information of the target object. Then, the underlying features are fused with the upsampled features by using a cascade aggregation module to generate a cascade feature map containing rich details and semantic information. At the same time, the boundary extraction module converts the underlying features into a boundary feature map, focusing on the edge information of the target object. Then, by performing boundary extension and semantic extension on the boundary feature map and the cascade feature map respectively, the boundary information and high-level semantic information in the feature map are further strengthened, and the quality of the feature map is further improved, so that the boundary final feature map and the cascade final feature map can more accurately represent the contour and internal structure of the target object. Finally, the target object is predicted in combination with the boundary final feature map and the cascade final feature map to obtain a segmentation result. The present invention effectively improves the accuracy of segmentation by comprehensively considering the boundary details and overall semantic information of the target object, especially when processing a target object with unclear boundaries or low contrast, the contour of the target object can be better identified and segmented. In a preferred embodiment of the present invention, the present invention can be used to segment polyps with unclear boundaries and low contrast in the medical field.

[0062] Optionally, extracting features from the detection image to obtain underlying features and up-sampled features includes:

[0063] Performing deep feature extraction on the detection image through an encoder network of a pyramid vision transformer to obtain a bottom-layer feature map and a deep-layer feature map;

[0064] The deep feature map is upsampled by an upsampling module of the encoder network to obtain an upsampled feature map with the same resolution as the original detection image;

[0065] The underlying feature map is combined with the upsampled feature map to obtain the underlying features and the corresponding upsampled features.

[0066] Specifically, the process of extracting features from the detection image to obtain underlying features and upsampled features is implemented by a pyramid visual transformer encoder network. Among them, the encoder network obtains underlying feature maps and deep feature maps from the detection image through its deep feature extraction capability, and these feature maps contain basic information and deep semantic information of the image. In this process, the convolutional layer and the pooling layer in the encoder network work together to reduce the spatial dimension of the detection image, while increasing the feature depth and extracting the semantic information of the image, so that the network can capture the key features in the image. Subsequently, through the upsampling module of the encoder network, the deep feature map is upsampled to the same resolution as the original detection image to generate an upsampled feature map. This step allows the network to restore the details of the deep feature map while maintaining spatial information, thereby obtaining a high-resolution feature representation. Finally, the underlying feature map is combined with the upsampled feature map to provide rich details and semantic information, laying a solid foundation for the subsequent cascade aggregation and boundary extraction modules. In a preferred embodiment of the present invention, the encoder network is used as a backbone network to extract multi-level long-distance dependency features, which are then used to construct a cascade aggregation module and a boundary extraction module.

[0067] In the embodiment of the present invention, by extracting features from the detection image, not only the network's ability to capture image details is enhanced, but also the understanding of deep semantic information of the image is improved. By reducing the spatial dimension and increasing the feature depth, the network can process image data more effectively and extract richer and more abstract feature representations.

[0068] Optionally, fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map includes:

[0069] The underlying features and the upsampled features are fused through the lightweight attention mechanism of the cascade aggregation module to obtain an attention-weighted feature map;

[0070] The underlying features are weighted from the channel dimension and the spatial dimension through the channel space attention module of the cascade aggregation module to obtain a channel attention feature map and a spatial attention feature map;

[0071] The cascade feature map is obtained according to the attention weighted feature map, the channel attention feature map and the spatial attention feature map.

[0072] Specifically, the role of the cascade aggregation module is to deeply fuse the underlying features and the upsampled features to generate a cascade feature map. This process involves a lightweight attention mechanism and a channel space attention module. The lightweight attention mechanism provides weighting for information transmission between different feature levels by calculating the correlation between the underlying features and the upsampled features, thereby realizing the extraction of high-level semantic information. In a preferred embodiment of the present invention, this mechanism includes two efficient attention mechanisms: lightweight attention and channel space attention modules, which are implemented by ReLU layers and Sigmoid activation functions as well as 1×1 convolution operations and batch normalization operations. Subsequently, the channel space attention module further weights the underlying features in terms of channel dimensions and spatial dimensions to obtain channel attention feature maps and spatial attention feature maps. This step is implemented by adaptive maximum pooling and adaptive average pooling functions and 1×1 convolution layers. Through lightweight attention and channel space attention modules, information transmission between different feature levels is promoted to realize the extraction of high-level semantic information. Finally, by combining the attention weighted feature map, channel attention feature map and spatial attention feature map, a cascade feature map that integrates multi-scale and multi-dimensional information is obtained, providing a rich feature basis for subsequent boundary extraction and morphological fusion.

[0073] In the embodiment of the present invention, by applying the lightweight attention mechanism and the channel space attention module, the network can pay more attention to the most informative part of the image, that is, the features that are most critical for the segmentation of the target object. This attention-driven feature fusion strategy not only enhances the ability of feature representation, but also improves the network's recognition accuracy of the target object contour, especially when dealing with target objects with low edge contrast or small size.

[0074] Optionally, the fusing of the underlying features and the upsampling through the lightweight attention mechanism of the cascade aggregation module to obtain an attention-weighted feature map includes:

[0075] Performing a convolution operation on the upsampled features and the underlying features of the detection image to obtain convolution operation results corresponding to the upsampled features and the underlying features respectively;

[0076] Performing normalization processing according to the convolution operation results corresponding to the up-sampled features and the underlying features respectively;

[0077] The normalized convolution operation result is subjected to light attention allocation through the ReLU layer activation function and the Sigmoid activation function in turn to obtain an attention-weighted feature map.

[0078] Specifically, first, the underlying features and upsampled features are processed by convolution operations respectively to obtain the convolution results of the underlying features and upsampled features, and then these convolution results are normalized, usually batch normalization, to reduce internal covariate shift and accelerate network training. Then, the normalized results pass through the ReLU layer activation function and the Sigmoid activation function in turn, where the ReLU activation function is used to introduce nonlinearity, and the Sigmoid activation function compresses the output value between 0 and 1 to achieve lightweight attention allocation. Finally, these attention-weighted feature maps are combined to form attention-weighted feature maps, which provide rich feature representations for subsequent segmentation tasks. In a preferred embodiment of the present invention, the two mechanisms of lightweight attention LA(·) and channel space attention CSAM(·) are specifically implemented as follows:

[0079]

[0080]

[0081] Among them, att(y, x): represents the output of the attention mechanism, where x and y are input features, and σ(·) correspond to the ReLU layer and the Sigmoid activation function, respectively. x (·), C y (·) and C s (·) represents a channel 1×1 convolution operation. BN(·) is a batch normalization operation. y and x are the upsampled features and underlying features from the encoder, respectively. C x (x) represents the channel 1×1 convolution operation on the input feature x, C y (y) represents the channel 1×1 convolution operation on the input feature y, C s (·) represents the channel 1×1 convolution operation on the output of the attention mechanism, and LA(y, x) represents the output of the lightweight attention.

[0082] In an embodiment of the present invention, through a lightweight attention mechanism, the network can pay more attention to the most important features for segmenting the target object, thereby improving the accuracy and efficiency of segmentation. This mechanism enables the network to dynamically adjust the attention to different features, and strengthens the recognition ability of the target object contour, especially when the target object contour is not obvious or the contrast is low. Through the application of ReLU and Sigmoid activation functions, the network can learn a more refined feature representation, which helps to better distinguish the target object from the background in the subsequent segmentation process.

[0083] Optionally, the channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including:

[0084] Through the channel space attention module, the underlying features are pooled using an adaptive maximum pooling function and an adaptive average pooling function;

[0085] The channel attention feature map is determined based on the underlying features after the pooling operation through the Sigmoid activation function and the Hadamard product.

[0086] Specifically, the role of the channel spatial attention module is to weight the underlying features from the channel dimension and the spatial dimension to obtain the channel attention feature map and the spatial attention feature map. This process first uses the adaptive maximum pooling function and the adaptive average pooling function to perform pooling operations on the underlying features to capture the global maximum and average values ​​in the feature map, respectively. These operations can extract important information in the feature map from different angles. Subsequently, the channel attention feature map is determined according to the results of the pooling operation through the Sigmoid activation function and the Hadamard product. In this way, the network can identify which channels contain the most important information for the segmentation task and give these channels higher weights. In a preferred embodiment of the present invention, the channel spatial attention module in the cascade aggregation module reduces the channel dimension by 16 times through a 1×1 convolution layer, followed by a ReLU activation function and another convolution operation to restore the channel dimension, thereby realizing the calculation of channel attention.

[0087] In a preferred embodiment of the present invention, the channel space attention module is used to perform pooling operations on the underlying features using an adaptive maximum pooling function and an adaptive average pooling function, which can be expressed as:

[0088]

[0089] CSAM(x)=CB(SA(CA(x)));

[0090] Among them, CA(x) represents the output of Channel Attention; SA(x) represents the output of Spatial Attention; CSAM(x) represents the output of Channel-Spatial AttentionModule; CB(·) represents the convolution block (ConvBlock), x is the result of the concat operation or the underlying feature F1 from the encoder, and "⊙" represents the Hadamard product. max (x) and Represent the adaptive maximum pooling and adaptive average pooling functions respectively. K i , i∈{1, 2} represents a convolutional layer with shared parameters, which consists of a channel-level 1×1 convolutional layer to reduce the channel dimension by a factor of 16, followed by a ReLU activation function and another 1×1 convolution operation to restore the channel dimension. and Respectively represent the maximum and average values ​​obtained along the channel dimension. Q(·) is a 7×7 convolutional layer with a padding value of 3. CAT(·): represents a connection operation, which is used to connect two feature maps along a certain dimension; represents the maximum value obtained along the channel dimension, represents the average value obtained along the channel dimension, and σ represents the activation function.

[0091] The F1, F2, and F3 features obtained by the PVT encoder are input into this cascade aggregation module to obtain the cascade feature map T2, and the cross-joint loss and cross-entropy loss between the true prediction and the feature map are used for optimization. The specific operation is as follows:

[0092]

[0093] Where UP(x) represents the output of the upsampling operation, Represents an upsampling operation with the scale factor set to 2. Convolution operation, used to transform channel dimensions. represents the batch normalization operation, Represents the ReLU activation function.

[0094]

[0095]

[0096]

[0097] Among them, CSAM represents the channel space attention module, which is used to process features; Represents the feature map after CSAM processing; is the connection operation along the channel dimension. LA represents the lightweight attention mechanism, which is used to process features. Finally, the 1×1 convolutional layer is used to process The sum is taken to obtain T2, which is the cascade feature map.

[0098] In an embodiment of the present invention, through the application of the channel space attention module, the network can more accurately identify and emphasize the most important features for target object segmentation. This mechanism enables the network to dynamically adjust the attention to different channels and spatial positions, and strengthens the ability to recognize the contour of the target object, especially when the contour of the target object is not obvious or the contrast is low.

[0099] Optionally, the channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including:

[0100] Through the channel space attention module, channel connection is performed according to the maximum value and average value obtained along the channel dimension of the underlying features;

[0101] The spatial attention feature map is determined by the Sigmoid activation function and the Hadamard product according to the underlying features after channel connection.

[0102] Specifically, the channel spatial attention module realizes channel connection by calculating the maximum and average values ​​of the underlying features along the channel dimension. This process involves adaptive maximum pooling and adaptive average pooling functions, which correspond to the maximum and average values ​​obtained along the channel dimension, respectively. Subsequently, these statistics are used for the channel connection operation to integrate the global channel statistics back into the feature map, enhancing the information expression ability of the feature map in the channel dimension. Then, through the Sigmoid activation function and the Hadamard product, the spatial attention feature map is determined according to the results after channel connection, so that the network can identify which spatial positions are more important for the segmentation of the target object and give these positions higher weights. In a preferred embodiment of the present invention, the channel spatial attention module reduces the channel dimension by 16 times through a 1×1 convolution layer, followed by a ReLU activation function and another convolution operation to restore the channel dimension, thereby realizing the calculation of channel attention.

[0103] In an embodiment of the present invention, the channel spatial attention module enables the network to focus more on the most important features for target object segmentation. By calculating the maximum and average values ​​in the channel dimension and performing channel connection, the network can capture global and representative features. Then, through the application of Sigmoid activation function and Hadamard product, the network can further refine these features and determine which specific spatial positions are more critical, thereby generating a spatial attention feature map. This mechanism not only enhances the ability of feature representation, but also improves the recognition accuracy of the target object contour, especially when dealing with target objects with low edge contrast or small size.

[0104] Optionally, the converting the underlying features into a boundary feature map by a boundary extraction module includes:

[0105] Inputting the bottom-level features into a preset convolutional layer of the boundary extraction module, and obtaining two adjacent feature cascades by reducing the channel size;

[0106] Inputting the features into a two-stream network in cascade order to obtain a plurality of boundary features, wherein each stream in the two-stream network uses a different convolution kernel;

[0107] The boundary features are combined by element-wise multiplication to obtain the boundary feature map.

[0108] Specifically, the process of converting the underlying features into a boundary feature map through the boundary extraction module is to capture the boundary information of the target object, which is crucial for the subsequent segmentation task. First, the underlying features are input into the preset convolutional layers of the boundary extraction module. These convolutional layers extract two adjacent feature cascades by reducing the channel size, which helps to capture boundary information from different scales. Next, these feature cascades are input into a two-stream network, and each stream uses a different convolution kernel. This design allows the network to capture boundary features from multiple angles and scales, enhancing the model's ability to recognize boundary details. Finally, these boundary features are combined through element-wise multiplication operations to obtain the final boundary feature map. This step enhances the model's ability to recognize and represent boundaries by fusing multi-scale boundary features. In a preferred embodiment of the present invention, a key component is introduced in the boundary extraction module: a camouflage removal block, which is used to remove camouflage around polyps. Specifically, F3 and F4′ are input into a 1×1 convolutional layer to reduce the channel size and obtain and F3 and F4′ represent the two feature maps input to the boundary extraction module. S3 and S4 represent the feature maps after being processed by the 1×1 convolution layer, their dimensions are reduced, and they have the same spatial resolution and number of channels.

[0109] The two features from adjacent layers are cascaded and input into a two-stream network, where each stream uses a different convolution kernel. This setting allows the streams to share information and capture boundary features across levels and multiple scales. The specific calculation process can be expressed as follows:

[0110]

[0111]

[0112]

[0113]

[0114] in, It is a continuous operation consisting of i×i convolutional layers, batch normalization, and ReLu activation functions. In order to fully integrate the multi-scale boundary features, the two features are multiplied using element-wise multiplication. and Put it together. and Represents the feature map after the first processing, in which 3×3 and 5×5 convolution kernels are used. and It represents the feature map after the second processing, which further integrates the multi-scale boundary features. Represents the connection operation along the channel dimension, which is used to merge features. Inspired by ResNet, the original cross-level features can be further combined through addition operations:

[0115]

[0116] in, and Represent element-wise product and addition, respectively. Subsequently, in order to further smooth the fused boundary features, it is input into a 3×3 convolutional layer to obtain the final boundary feature map Among them, B fused Represents the fused feature map, which is obtained by element-wise multiplication and addition operations.

[0117] In an embodiment of the present invention, through the processing of the boundary extraction module, the network can more accurately identify and represent the boundary of the target object. The operation of reducing the channel size helps to extract more refined boundary features, and the design of the two-stream network enables the model to capture boundary information from different scales and angles.

[0118] Optionally, performing boundary expansion and semantic expansion on the boundary feature map and the cascade feature map respectively to obtain a boundary final feature map and a cascade final feature map includes:

[0119] Separating the boundary region and the internal region of the target object from the cascade feature map by using a morphological operator;

[0120] Fusing the boundary region and the internal region of the object to obtain a final feature map of the boundary;

[0121] Extracting high-level semantic features from the cascaded feature maps;

[0122] The cascade feature map is expanded according to the high-level semantic features through a cross non-local network to obtain the cascade final feature map.

[0123] Specifically, the process of boundary extension and semantic extension of the boundary feature map and the cascade feature map is to enhance the feature representation of the boundary and internal area of ​​the target object and to improve the expression of high-level semantic information. First, the boundary area and internal area of ​​the target object are separated from the cascade feature map by morphological operators, especially erosion operators and dilation operators. Then, the separated boundary area and internal area are fused to obtain the boundary final feature map. This fusion process helps to retain both boundary details and internal structure information. Then, high-level semantic features are extracted from the cascade feature map. These features contain rich contextual information, which is crucial for understanding the overall structure of the target object. Finally, the cascade feature map is extended according to the extracted high-level semantic features through a cross non-local network to obtain a cascade final feature map, which illustrates the use of the Transformer attention mechanism to enhance the semantic information of the feature map. In a preferred embodiment of the present invention, the boundary and internal information are separated from the feature map according to the morphological operators (dilation and erosion) through a morphological fusion module, and then the information is fused through a cross non-local network to expand to the entire target object area.

[0124] In a preferred embodiment of the present invention, in order to explore the intrinsic relationship between the low-level boundary features output by the boundary extraction module and the high-level semantic features output by the cascade aggregation module, a novel cross non-local network and morphological operators are introduced to implement the morphological fusion module. Two morphological operators are introduced to separate the boundary and internal polyp regions from the feature map T2. Specifically, given T2 (the feature map output by the cascade aggregation module), the channel dimension is compressed to 1 by 1×1 convolution to obtain the initial segmentation mask f mask Then, using the erosion operator (ε) and the dilation operator To separate the boundary and interior regions:

[0125]

[0126]

[0127] TB i =TB×T i ;

[0128] Among them, f mask represents the initial segmentation mask, which is obtained from T2 through 1×1 convolution, and the channel dimension is compressed to 1; ε represents the erosion operator, which is used for morphological operations to separate the boundary and internal areas; represents the dilation operator, which is used for morphological operations to separate the boundary and internal regions; TR is the internal region of T, and TB represents the potential boundary region of T. Here, is the number of operations. iAdd TB to get TBi. TBi represents the potential boundary area considering the i-th feature. i Multiply element by element to get, T i represents the i-th feature map, which is used to combine with TB to obtain TB i Different from the previous choice of non-local network, considering the finiteness of linear function expression, g(·) is not simply set to linear: g(x j )=W g x j ,(W g is the weight matrix, x j is the input feature) but is replaced by an efficient Transformer attention mechanism:

[0129]

[0130]

[0131]

[0132] where a l (x) represents a smaller-scale local branch processing, a g (x) Larger-scale global branch processing, G(x) represents the final output, which is obtained by combining the results of local and global branch processing; represents an efficient Transformer block with a block size of p×p for obtaining local element-wise attention maps, and p is set to 4 for best performance. represents an efficient Transformer block at the entire feature map resolution, used to obtain the global element attention map; σ represents the activation function.

[0133] Given TB i and T j (i∈{1,2},j∈{1,2},i+j=3), where TB i represents the i-th potential boundary region feature map, T j represents the jth feature map, and fuses them through self-attention; first, two linear functions, a convolution operation with a kernel size of 1×1, are performed on T j Perform W K (·) and W V (·) Dimensionality reduction to obtain feature maps and For TB i , similarly, using the convolutional layer W Q (·) to obtain

[0134] K j =WK (T j ), V j =W V (T j ), Q i =W Q (TB i );

[0135] Among them, K j W K (·) Convolutional layers from T j The obtained feature map is used to calculate the key (Key), V j Indicates that through W V (·) Convolutional layers from T j The obtained feature map is used to calculate the value (Value), Q i Indicates that through W Q (·) Convolutional layers from TB i The obtained feature map is used to calculate the query.

[0136] Apply the softmax function on the channel dimension and select the second channel as the attention map, denoted as F(·), where F(·) represents the attention map obtained by applying the softmax function on the channel dimension; get in, K j and The result of the Hadamard product (element multiplication) between them, then calculate K j and The Hadamard product between them gives This product is K j Different pixels of the feature are assigned different weights. Subsequently, an adaptive pooling AP(·) operation is used to reduce the displacement of the feature. AP(·) represents the adaptive pooling operation, which is used to reduce the displacement of the feature and perform center cropping to obtain the feature map f j ∈R 4×4×16 , f j represents the feature map obtained after the adaptive pooling operation. Then, Q is established by the inner product j and K j The correlation between each pixel in . Specifically, this process can be written as follows:

[0137]

[0138]

[0139] Where AP(·) represents the adaptive pooling operation, f i T Yes iThe transpose of is the attention map. Then, V j The feature map captures global attention through G(·), and we get Afterwards, it is combined with the attention map Multiply them together and input the resulting features into the graph convolution layer GCN(·) to get g i ∈R 4×4×16 Then, calculate and g j The inner product between:

[0140]

[0141] Where G represents the global attention feature map obtained by G(·), GCN(·) represents the graph convolution layer used to process the feature map, and g i Represents the feature map obtained after processing by the graph convolution layer.

[0142] Finally, a 1×1 convolutional layer W is used j Y j Adjust to T j The same channel dimension, then with T j Add together and get The process is as follows:

[0143] out(j+2)=T j +W j (Y j ).

[0144] Among them, W j Represents a 1×1 convolutional layer used to adjust the feature map Y j The channel dimension of Y j It means that the feature map after a series of processing needs to be adjusted to be consistent with T j Same channel dimension; T j Represents the original feature map and the adjusted Y j Add; out(j+2) represents the final output feature map, which is T j and through W j Adjusted Y j The result of addition; Represents the dimension of the output feature map, where H and W are the height and width of the feature map respectively, and 32 is the number of channels.

[0145] In the embodiment of the present invention, the boundary and internal area are effectively separated and fused by morphological operators, which improves the clarity of boundary features and the integrity of internal structures. At the same time, the extraction of high-level semantic features and the application of cross-non-local networks enable the network to better understand the contextual information of the target object and improve the semantic expression ability of the feature map.

[0146] Optionally, predicting the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object includes:

[0147] Inputting the boundary final feature map and the cascade final feature map into a preset convolutional layer, predicting the target object through the preset convolutional layer, and obtaining prediction results corresponding to the boundary final feature map and the cascade final feature map respectively;

[0148] The prediction results corresponding to the boundary final feature map and the cascade final feature map are summed to obtain the segmentation result of the target object.

[0149] Specifically, the process of predicting the segmentation result of the target object is achieved by inputting the boundary final feature map and the cascade final feature map into the preset convolution layer. The outputs of the cascade aggregation module and the boundary extraction module are used to provide rich boundary information and high-level semantic information, respectively. The preset convolution layer processes these feature maps to generate their corresponding prediction results. This process involves upsampling operations and convolution operations for converting channel dimensions, as well as connection operations along the channel dimensions. Finally, the segmentation result of the target object is obtained by summing the prediction results corresponding to the boundary final feature map and the cascade final feature map. This step reflects the advantages of feature fusion and integrates information of different dimensions to improve the accuracy of segmentation.

[0150] In a preferred embodiment of the present invention, the prediction segmentation loss is expressed as:

[0151] P = (out3 + T2) + (out4 + T2);

[0152]

[0153] Among them, P represents the final predicted segmentation result, out3 and out4 represent the two feature maps output by the model, and T2 represents the feature map output by the cascade aggregation module. It is the segmentation loss, which calculates the difference between the final segmentation result P and the true value G. represents the weighted intersection-over-union (IoU) loss, Denotes; weighted binary cross entropy (BCE) loss. G 表示真实值。 These losses constrain the prediction graph from both object-level (global structure) and pixel-level (local details) perspectives. Unlike the standard BCE loss that treats all pixels equally, Assign higher weights to difficult pixels. Similarly, compared with the standard IoU loss, Give more importance to difficult pixels.

[0154] It is worth mentioning that the above predictions need to be trained in advance. Typically, this network can be implemented on an RTX4090 GPU with 24GB of memory using the PyTorch framework. The network is updated and optimized with a batch size of 16. The backbone of the proposed network is the encoder network. Since the sizes of polyp images vary, a multi-scale strategy is adopted during training. The hyperparameters are set as follows: Adam optimizer is used with a learning rate of 1×10 -4 , the weight decay is 1×10 -4 Inspired by PraNet, the input images are resized to 352×352 for 100 epochs. During training, the images are scaled by factors of 0.75, 1, and 1.25 instead of using data augmentation.

[0155] In the embodiment of the present invention, by combining the prediction results of the boundary final feature map and the cascade final feature map, the network can obtain a segmentation result that integrates boundary details and high-level semantic information. This fusion strategy enables the segmentation result to not only accurately capture the contour of the target object, but also understand the overall structure and context of the object, thereby improving the accuracy and robustness of the segmentation.

[0156] In summary, in a preferred embodiment of the present invention, for the polyp segmentation of medical images, the target object segmentation method of the present invention is adopted. First, given a colonoscopy image I∈R H×W×3 , using the Pyramid Visual Transformer (PVT) encoder network as the backbone to extract multi-level long-distance dependency features Among them, C i ∈{512,320,128,64} and i∈{1,2,3,4}. Features F1, F2, and F3 are input into CAM for fusion to obtain feature maps Similarly, Resized to The low-level features F3 and F4′ are transformed into Next, T1 and T2 are aligned and fused through MFM. Morphological operators (dilation and erosion) are first applied to separate the boundary and internal information from T1 and T2. Then, these processed features are fused through a novel cross non-local network to expand the boundary and high-level semantic information to the entire polyp area, resulting in feature maps out3 and out4. Finally, out3+T2 and out4+T2 are passed through a 1×1 convolutional layer to predict the polyp segmentation results P1, P2. The sum of P1 and P2 is taken as the final segmentation result.

[0157] In order to verify the effectiveness of the target object segmentation method, polyps with unclear structural boundaries in the medical field were taken as test objects. The performance of the PolypSolver method was tested on five polyp image datasets (Kvasir-SEG, ClinicDB, ColonDB, Endoscene, and ETIS), and compared with nine currently popular polyp segmentation methods.

[0158] The information of the five polyp image datasets is as follows: (1) Kvasir-SEG: Contains 1,000 high-quality endoscopic images and corresponding polyp segmentation masks. The image resolution is 1920×1080 or 3888×2916 pixels. The images cover various polyp types and shapes and are diverse. (2) ClinicDB: Contains 612 endoscopic images. These images have been manually annotated and the location and size of polyps are marked. The image resolution is 288×384 pixels and contains information on the shape, size and color of various polyps. (3) ColonDB: Contains 380 endoscopic images with an image resolution of 574×500 pixels. The images cover a variety of polyp morphologies and have good diversity. They provide accurate segmentation masks for polyps. (4) Endoscene: Contains 60 video clips and a total of 300 images. These images are all captured from endoscopic videos and provide pixel-level segmentation masks for polyps. The image quality and resolution are high and suitable for training complex deep learning models. The dataset contains dynamic scenes and is suitable for video-level detection and segmentation tasks. (5) ETIS: Contains 196 endoscopic images with an image resolution of 1225×966 pixels. The image quality is high and shows different sizes, shapes, and textures of polyps. The dataset contains a variety of polyp manifestations, and the proportion of polyps in the images varies greatly.

[0159] The information of the 9 polyp segmentation methods is as follows: (1) UNet: This method ensures the feature alignment of the encoder and decoder parts through a symmetrical U-shaped architecture and jump connections, reducing the loss of feature information in the deep layer of the network. (2) UNet++: This method uses enhanced jump connections and hierarchical feature optimization to more fully share and combine features in network layers of different depths, and can better learn complex segmentation boundaries and details. (3) PraNet: This method combines high-level features with parallel partial decoders and introduces a reverse attention network to capture boundary clues. (4) SANet: This method uses a shallow attention module to filter out background noise in shallow features. (5) Polyp-PVT: This method introduces a camouflage recognition module to extract boundary information and a similarity aggregation module to expand the high-level semantic location information of the polyp area. (6) PVT-Cascade: This method uses the multi-stage feature representation of the visual transformer to simultaneously learn multi-scale and multi-resolution spatial representations to achieve faster convergence of model accuracy and improve decoder performance. (7) M2SNet: This method equips different levels of multi-scale SU with different receptive fields in a pyramid form, thereby achieving inter-layer multi-scale feature aggregation and obtaining rich multi-scale difference information. (8) CFA-Net: This method designs a segmentation network based on a two-stream structure to exploit the hierarchical semantic information in cross-layer features. (9) NPD-Net: This method proposes a novel deep supervision strategy that directly utilizes the GT encoder (encoding the GT to obtain its graph) after initialization to alleviate overfitting and enhance generalization ability without pre-training.

[0160] In this experiment, the training set includes 900 images from Kvasir-SEG and 550 images from ClinicDB. The test set includes 100 images from Kvasir-SEG, 62 images from CVC-ClincDB, 380 images from CVC-ColonDB, 60 images from Endoscene, and 196 images from ETIS.

[0161] After testing, the present invention is superior to the existing methods. On the ClinicDB dataset, the mDic score of our model is only 0.1% lower than Polyp-PVT and 2.0% higher than SANet. On the Kvasir-SEG dataset, the mDic score of our model is 0.3% higher than Polyp-PVT and 2.2% higher than PraNet.

[0162] Combination Figure 3 As shown, the present invention also provides a target object segmentation system, comprising:

[0163] An acquisition unit is used to acquire a detection image containing a target object, and perform feature extraction on the detection image to obtain underlying features and up-sampled features;

[0164] A fusion unit, used for fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map;

[0165] An expansion unit, configured to convert the underlying features into a boundary feature map through a boundary extraction module and perform boundary expansion and semantic expansion on the boundary feature map and the cascade feature map, respectively, to obtain a boundary final feature map and a cascade final feature map;

[0166] A prediction unit is used to predict the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

[0167] The target object segmentation system of the present invention obtains a detection image containing a target object and extracts underlying features and upsampled features from it, and these features include the shape, texture and context information of the target object. Then, the underlying features are fused with the upsampled features by using a cascade aggregation module to generate a cascade feature map containing rich details and semantic information. At the same time, the boundary extraction module converts the underlying features into a boundary feature map, focusing on the edge information of the target object. Then, by performing boundary extension and semantic extension on the boundary feature map and the cascade feature map respectively, the boundary information and high-level semantic information in the feature map are further strengthened, and the quality of the feature map is further improved, so that the boundary final feature map and the cascade final feature map can more accurately represent the contour and internal structure of the target object. Finally, the target object is predicted in combination with the boundary final feature map and the cascade final feature map to obtain a segmentation result. The present invention effectively improves the accuracy of segmentation by comprehensively considering the boundary details and overall semantic information of the target object, especially when processing a target object with unclear boundaries or low contrast, it can better identify and segment the contour of the target object, such as segmenting polyps with unclear boundaries and low contrast in the medical field.

[0168] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0169] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0170] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those 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 claimed herein.

Claims

1. A target object segmentation method, characterized in that: include: Acquire a detection image containing a target object, and perform feature extraction on the detection image to obtain underlying features and up-sampled features; The bottom-level features and the up-sampled features are fused through a cascade aggregation module to obtain a cascade feature map; The bottom layer features are converted into a boundary feature map by a boundary extraction module, and the boundary feature map and the cascade feature map are respectively subjected to boundary expansion and semantic expansion to obtain a boundary final feature map and a cascade final feature map; The target object is predicted according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

2. The target object segmentation method according to claim 1, characterized in that: The step of extracting features from the detection image to obtain underlying features and up-sampled features includes: Performing deep feature extraction on the detection image through an encoder network of a pyramid vision transformer to obtain a bottom-layer feature map and a deep-layer feature map; The deep feature map is upsampled by an upsampling module of the encoder network to obtain an upsampled feature map with the same resolution as the original detection image; The underlying feature map is combined with the upsampled feature map to obtain the underlying features and the corresponding upsampled features.

3. The target object segmentation method according to claim 2, characterized in that: The step of fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map includes: The underlying features and the upsampled features are fused through the lightweight attention mechanism of the cascade aggregation module to obtain an attention-weighted feature map; The underlying features are weighted from the channel dimension and the spatial dimension through the channel space attention module of the cascade aggregation module to obtain a channel attention feature map and a spatial attention feature map; The cascade feature map is obtained according to the attention weighted feature map, the channel attention feature map and the spatial attention feature map.

4. The target object segmentation method according to claim 3, characterized in that: The light attention mechanism of the cascade aggregation module is used to fuse the underlying features and the upsampling to obtain an attention-weighted feature map, including: Performing a convolution operation on the upsampled features and the underlying features of the detection image to obtain convolution operation results corresponding to the upsampled features and the underlying features respectively; Performing normalization processing according to the convolution operation results corresponding to the up-sampled features and the underlying features respectively; The normalized convolution operation result is subjected to light attention allocation through the ReLU layer activation function and the Sigmoid activation function in turn to obtain an attention-weighted feature map.

5. The target object segmentation method according to claim 3, characterized in that: The channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including: Through the channel space attention module, the underlying features are pooled using an adaptive maximum pooling function and an adaptive average pooling function; The channel attention feature map is determined based on the underlying features after the pooling operation through the Sigmoid activation function and the Hadamard product.

6. The target object segmentation method according to claim 3, characterized in that: The channel spatial attention module of the cascade aggregation module weights the underlying features from the channel dimension and the spatial dimension to obtain a channel attention feature map and a spatial attention feature map, including: Through the channel space attention module, channel connection is performed according to the maximum value and average value obtained along the channel dimension of the underlying features; The spatial attention feature map is determined by the Sigmoid activation function and the Hadamard product according to the underlying features after channel connection.

7. The target object segmentation method according to claim 1, characterized in that: The step of converting the underlying features into a boundary feature map by a boundary extraction module includes: Inputting the bottom-level features into a preset convolutional layer of the boundary extraction module, and obtaining two adjacent feature cascades by reducing the channel size; Inputting the features into a two-stream network in cascade order to obtain a plurality of boundary features, wherein each stream in the two-stream network uses a different convolution kernel; The boundary features are combined by element-wise multiplication to obtain the boundary feature map.

8. The target object segmentation method according to claim 1, characterized in that: The step of performing boundary expansion and semantic expansion on the boundary feature map and the cascade feature map respectively to obtain a boundary final feature map and a cascade final feature map comprises: Separating the boundary region and the internal region of the target object from the cascade feature map by using a morphological operator; Fusing the boundary region and the internal region of the object to obtain a final feature map of the boundary; Extracting high-level semantic features from the cascaded feature maps; The cascade feature map is expanded according to the high-level semantic features through a cross non-local network to obtain the cascade final feature map.

9. The target object segmentation method according to claim 1, characterized in that: The step of predicting the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object includes: Inputting the boundary final feature map and the cascade final feature map into a preset convolutional layer, predicting the target object through the preset convolutional layer, and obtaining prediction results corresponding to the boundary final feature map and the cascade final feature map respectively; The prediction results corresponding to the boundary final feature map and the cascade final feature map are summed to obtain the segmentation result of the target object.

10. A target object segmentation system, characterized in that: include: An acquisition unit is used to acquire a detection image containing a target object, and perform feature extraction on the detection image to obtain underlying features and up-sampled features; A fusion unit, used for fusing the underlying features and the upsampled features through a cascade aggregation module to obtain a cascade feature map; An expansion unit, configured to convert the underlying features into a boundary feature map through a boundary extraction module and perform boundary expansion and semantic expansion on the boundary feature map and the cascade feature map, respectively, to obtain a boundary final feature map and a cascade final feature map; A prediction unit is used to predict the target object according to the boundary final feature map and the cascade final feature map to obtain a segmentation result of the target object.

Images