Target segmentation method for boundary learning optimization

Abstract

The invention discloses a target segmentation method for boundary learning optimization, and relates to the field of image segmentation, and the method comprises the steps: constructing and training a target segmentation model; inputting the airport airspace image into the trained target segmentation model to obtain a prediction mask, and completing the segmentation of the target in the airport airspace; wherein the airport airspace image is input into the trained target segmentation model to obtain the prediction mask, specifically, the airport airspace image is input into the multi-scale feature extraction module to obtain features of different scales; inputting the features of different scales into a deformable convolutional coding module to obtain a plurality of enhanced features; inputting the plurality of enhanced multi-scale features into a multi-scale feature fusion module to obtain fusion features; according to the method and the device, the segmentation accuracy of the target segmentation model can be improved, and more local detail information is reserved, so that the pixel quality of the segmented airplanes in the airport airspace is improved.

Description

Technical Field

[0001] The present application relates to the field of image segmentation, and in particular to a target segmentation method based on boundary learning optimization. Background Art

[0002] In the field of computer vision and remote sensing image analysis, accurate segmentation of sea and air targets such as aircraft and ships is a key task, which is of great significance to many application scenarios such as air traffic management, civil aviation safety, and maritime traffic management.

[0003] In some cases, rule-based object segmentation methods often suffer from poor generalization and adaptability when faced with complex and changing environments. In recent years, deep learning methods, particularly convolutional neural networks, have achieved remarkable success in image segmentation tasks thanks to their superior feature extraction capabilities and end-to-end learning framework.

[0004] However, most existing deep learning models rely on pixel-level loss functions, such as cross-entropy loss and Dice loss. These algorithms still suffer from some drawbacks when processing complex objects: pixel-level loss functions can easily lead to blurred boundaries in segmentation results or loss of detailed information, thus limiting further improvement in algorithm performance. This loss of detail and blurred boundaries are particularly prominent when segmenting objects with fine edges and complex shapes, such as aircraft and ships. Summary of the Invention

[0005] The purpose of this application is to provide a target segmentation method with boundary learning optimization, which can improve the accuracy of target segmentation model segmentation, retain more local detail information, and thus improve the pixel quality of aircraft segmented in airport airspace.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a target segmentation method for boundary learning optimization, including the following contents.

[0008] Construct and train a target segmentation model; the loss function used for training the target segmentation model is constructed based on a pixel-level loss function, a Fourier descriptor loss function and an adaptive weighting coefficient; the target segmentation model includes: a multi-scale feature extraction module, a deformable convolutional coding module, a multi-scale feature fusion module and a mask generation module connected in sequence.

[0009] The airport airspace image is input into the trained target segmentation model to obtain a prediction mask to complete the segmentation of the targets in the airport airspace; the prediction mask is the pixel position of the aircraft in the airport airspace image.

[0010] Among them, the airport airspace image is input into the trained target segmentation model to obtain the prediction mask, which specifically includes the following contents.

[0011] The airport airspace image is input into the multi-scale feature extraction module to obtain features of different scales.

[0012] Features of different scales are input into the deformable convolutional coding module to obtain multiple enhanced features.

[0013] The multiple enhanced feature multi-scale features are input into the multi-scale feature fusion module to obtain fused features.

[0014] The fused features are input into the mask generation module to obtain a predicted mask.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects.

[0016] (1) A loss function for training the target segmentation model is constructed by using pixel-level loss function, Fourier descriptor loss function and adaptive weighting coefficient. Adding the Fourier descriptor loss function on the basis of the pixel-level loss function can effectively capture the topological structure of the shape, prompting the model to output a prediction mask that is not only close to the true value at the pixel level, but also more similar in shape features, retaining more details, making the boundaries clearer, and improving the prediction effect of the final prediction mask.

[0017] (2) The deformable convolutional coding module, on the one hand, can prompt the convolution kernel to dynamically adjust the sampling position according to the input features, so as to better adapt to the shape changes of the target at different rotation angles and different shooting angles, and more accurately capture the edges and details of these small targets to avoid omission or mis-segmentation; on the other hand, the deformable convolutional coding module can capture global context information through the self-attention mechanism, so that the model can not only pay attention to local features, but also understand the overall structure and contextual relationship of the target, further enhancing the ability of feature representation; thirdly, by designing sparse attention selection, the number of sampling points of the deformable convolution can be reduced; at the same time, the combination of sparse deformable convolution and learnable gating unit can strengthen the features of key channels and spatial positions, reduce the computational complexity of invalid features, and improve the accuracy of target segmentation model segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 A flowchart of a target segmentation method for boundary learning optimization provided in one embodiment of the present application.

[0020] Figure 2 for Figure 1 Schematic diagram of the refinement process of an object segmentation method with boundary learning optimization. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0023] In an exemplary embodiment, Figure 1 and Figure 2 As shown, a target segmentation method for boundary learning optimization is provided. The method is executed by a computer device, and can be executed separately by a computer device such as a terminal or a server, or can be executed jointly by a terminal and a server. In the embodiment of the present application, the method is applied to a server as an example for explanation, including the following steps S1 to S2.

[0024] Step S1: Construct and train a target segmentation model; the loss function used for training the target segmentation model is constructed based on a pixel-level loss function, a Fourier descriptor loss function, and an adaptive weighting coefficient; the target segmentation model includes: a multi-scale feature extraction module, a deformable convolutional encoding module, a multi-scale feature fusion module, and a mask generation module connected in sequence.

[0025] Furthermore, the deformable convolutional coding module includes: a channel-space separation offset prediction unit, a sparse attention selection unit, a deformable convolution unit, a learnable gating unit and a bidirectional heterogeneous feature interaction unit.

[0026] Furthermore, the learnable gating unit includes: a channel attention branch and a spatial attention branch.

[0027] Specifically, the mask generation module includes three deconvolution layers and a sigmoid layer connected in sequence.

[0028] Furthermore, the training process of the target segmentation model specifically includes: constructing a sample data pair; the sample data pair includes: a sample airport airspace image and a sample true mask; inputting the sample airport airspace image into the target segmentation model to obtain a sample prediction mask; based on the sample prediction mask and the corresponding sample true mask, using a loss function to iteratively optimize the parameters of the target segmentation model until the loss function reaches a minimum value or the iterative optimization round reaches a maximum value, stopping the iterative optimization, and obtaining a trained target segmentation model.

[0029] Furthermore, the expression of the loss function is as follows.

[0030] .

[0031] .

[0032] .

[0033] .

[0034] in, is the loss function; is the adaptive weighting coefficient; is the pixel-level loss function; is the Fourier descriptor loss function; Predict the mask for the sample pixel values; is the true mask of the sample pixel values; is the total number of pixels; is the total number of Fourier descriptors; is the true mask of the sample Fourier descriptors; is the 0th Fourier descriptor of the true mask of the sample; The first prediction mask for the sample Fourier descriptors; The 0th Fourier descriptor of the sample prediction mask; To take the model; is the number of current iteration training; is the total number of training iterations.

[0035] Specifically, before training begins, images of the airport airspace (or images of other scenes with objects) are captured using the front camera of the air tower. These images are then cropped and resized to 512×512×3. Masks are then used to annotate the locations of aircraft objects, creating sample data pairs: <sample airport airspace image, sample ground-truth mask>. Pixels in the airport airspace image that belong to aircraft objects are labeled as 1 in the ground-truth mask, while pixels that do not belong to aircraft objects are labeled as 0 in the ground-truth mask. All data pairs are then divided into training and test sets with a 7:3 ratio.

[0036] During training, repeat the following steps 500 times: data pairs, and input the images into the target segmentation model respectively to obtain The prediction result is the sample prediction mask. Then the total loss function corresponding to each data pair is obtained , based on Total loss function The average value of is returned to optimize the network parameters in the target segmentation model. In this embodiment, .

[0037] (1) The pixel-level loss function is calculated as follows.

[0038] .

[0039] (2) The Fourier descriptor loss function is calculated as follows.

[0040] First, construct the edge point coordinate sequences of the sample real mask and the sample predicted mask.

[0041] The definition of an edge point is: if a pixel value in the mask is 1 and the values ​​of the 8 adjacent pixels are not all 1, then the pixel is an edge point.

[0042] For any mask, starting from the edge point with the smallest X coordinate value, take edge points in a clockwise direction to obtain the edge point sequence corresponding to the mask. The edge point sequence of the sample's true mask is expressed as ,in is the first true mask of the sample The X and Y coordinates of edge points; the edge point sequence of the sample prediction mask is expressed as ,in is the first The X and Y coordinates of each edge point.

[0043] Then, the Fourier descriptors of the true mask and the predicted mask are obtained as follows.

[0044] .

[0045] .

[0046] in, 、 are the first Fourier descriptor and sample prediction mask Fourier descriptors, ; is the total number of Fourier descriptors. is the number of edge points in the edge point sequence of the sample's true mask, is the number of edge points in the edge point sequence of the sample prediction mask.

[0047] The low-frequency part of the Fourier descriptor mainly describes the overall structure of the shape, while the high-frequency part describes the details. In order to reduce the amount of calculation and focus on the main features of the shape, let .

[0048] Finally, calculate the Fourier descriptor loss function .

[0049] .

[0050] The above formula is normalized by dividing the descriptor by the first Fourier descriptor used to represent the centroid position and total length of the contour, making the Fourier descriptor insensitive to scale changes.

[0051] Designing a Fourier Descriptor-based loss function and combining it with the MSE loss function effectively captures shape information. A Fourier descriptor is a set of coefficients obtained by applying a Fourier transform to the complex representation of a contour. It describes the key features of a shape and is invariant to translation, rotation, and scaling. By incorporating the Fourier descriptor loss into the loss function, the predicted masks generated by the model are not only closer to the true values ​​at the pixel level, but also more similar in shape characteristics.

[0052] (3) The weighting coefficient is determined as follows.

[0053] Adaptive weighting coefficient Used to balance the importance of the two losses. In order to better adjust the training process, set the adaptive weighting coefficient The value is set as follows.

[0054] .

[0055] After training is completed, the model is tested using the test set to evaluate its segmentation performance.

[0056] Step S2: Transform the airport airspace image (i.e. Figure 2 The image to be segmented) is input into the trained target segmentation model to obtain the predicted mask (i.e. Figure 2 The predicted mask of the image to be segmented is used to complete the segmentation of the target in the airport airspace; the predicted mask is the pixel position of the aircraft in the airport airspace image.

[0057] The airport airspace image is input into the trained target segmentation model to obtain a prediction mask, which specifically includes steps S21 to S24.

[0058] Step S21: Input the airport airspace image into the multi-scale feature extraction module to obtain features of different scales.

[0059] Specifically, in this embodiment, the multi-scale feature extraction module extracts 、 、 There are three features, and the shapes of the three are 128×128×64, 64×64×128, and 32×32×256 respectively.

[0060] Step S22: Input the features of different scales into the deformable convolutional coding module to obtain multiple enhanced features.

[0061] Specifically, a common deformable convolutional coding module is the Deformable DETR (DETR, short for Detection Transformer, a Transformer-based end-to-end object detection network). By introducing a deformable attention module and a multi-scale feature fusion mechanism, it can effectively improve feature learning efficiency and accuracy. However, this module typically requires a large number of sampling points, resulting in high computational redundancy. Therefore, the Deformable DETR, combined with a dynamic sparse network, achieves high computational efficiency and better learning results with a smaller number of sampling points.

[0062] The input of the deformable convolutional coding module is one of the features output by the multi-scale feature extraction module , is the feature number, and the output is the corresponding enhanced feature .feature Number of channels and enhanced features The number of channels are and .

[0063] The deformable convolutional coding module includes a channel-space separation offset prediction unit, a sparse attention selection unit, a deformable convolution unit, a learnable gating unit, and a bidirectional heterogeneous feature interaction unit. The channel-space separation offset prediction unit is used to Perform channel-space separation offset prediction to obtain the spatial offset and channel modulation weights Sparse attention selection unit to channel modulation weight Select the weight elements in to get the sparse attention weight The deformable convolution unit is based on the spatial offset and the attention weight after sparsification , for the input features Perform deformable convolution to obtain feature maps The learnable gating unit includes a channel attention branch and a spatial attention branch, both of which are based on feature maps. Get the feature maps after channel attention weighting respectively And the feature map after spatial attention weighting The bidirectional heterogeneous feature interaction unit weights the feature map after channel attention And the feature map after spatial attention weighting Fusion to obtain enhanced features .

[0064] Furthermore, step S22 specifically includes steps S221 to S225.

[0065] Step S221: Input features of different scales into a channel-space separation offset prediction unit to obtain spatial offsets and channel modulation weights corresponding to features of different scales.

[0066] Specifically, based on the input features Perform channel-space separation offset prediction to obtain the spatial offset and channel modulation weights : ,in Indicates the channel-space separation offset prediction method. Spatial offset , preset parameters Related to the number of sampling points of deformable convolution, in this embodiment, 、 、 Corresponding They are 64, 128, and 256 respectively. , Based on characteristics Number of channels and Enhanced Features Number of channels The number of weights set must match the feature channel transformation requirements. In this embodiment, to ensure that the channel modulation weights can accurately act on each original feature channel, set .

[0067] Step S222: Input the channel modulation weight into the sparse attention selection unit for weight selection to obtain the sparse attention weight.

[0068] Furthermore, step S222 specifically includes steps S2221 to S2225.

[0069] Step S2221: Calculate the channel modulation weight score based on the channel modulation weight.

[0070] Furthermore, the calculation formula of the channel modulation weight score is as follows.

[0071] .

[0072] in, Score the channel modulation weight; is the channel modulation weight; For noise; is the temperature parameter.

[0073] Step S2222: Select the channel modulation weight score and determine the index.

[0074] Furthermore, the expression of the index is as follows.

[0075] .

[0076] in, is the index; For selection operation; Score the channel modulation weight; is the number of importance weights chosen.

[0077] Step S2223: Construct a sparse selection matrix based on the index.

[0078] Step S2224: Based on the sparse selection matrix and channel modulation weights, the sparse attention weights are obtained.

[0079] Furthermore, the expression of the attention weight after sparsification is as follows.

[0080] .

[0081] in, is the attention weight after sparsification; Select the matrix for sparseness; is the dot product; is the channel modulation weight.

[0082] Specifically, Gumbel noise is generated first , then calculate the channel modulation weight score , and then based on the channel modulation weight score Get the index through Top-k operation . Then construct the sparse selection matrix , sparse selection matrix Corresponding index The element value of is 1, and the rest are 0. Finally, the sparse attention weight is obtained .

[0083] Step S223: Input the scale feature, spatial offset and sparse attention weight into the deformable convolution unit for convolution to obtain a feature map.

[0084] Furthermore, the expression of the feature map is as follows.

[0085] .

[0086] in, is the feature map; is the convolution operation of the deformable convolution unit, is a deformable convolution kernel, is the side length of the sampling window (e.g. Convolution kernel window corresponds to ), is the total number of sampling points; For the scale characteristics; is a deformable convolution kernel; is the attention weight after sparsification; is the spatial offset; for The field of real numbers, is the height of the feature map; is the width of the feature map; is the number of output channels.

[0087] Specifically, the attention weight after sparseness Used to assign dynamic weighting coefficients to each sampling point to enhance the feature contribution of key sampling positions; spatial offset It is used to determine the displacement value of the coordinates of each sampling point during convolution, and to make the sampling points deviate from the regular grid through learning, so as to adapt to the target shape.

[0088] Step S224: Input the feature map into the learnable gating unit to obtain the channel attention weighted feature map and the spatial attention weighted feature map.

[0089] Specifically, the processing process of the channel attention branch is: first, the feature map Perform global average pooling to obtain the channel description vector , and then pass through the first fully connected layer and the second fully connected layer to obtain the channel attention weight The weight of the first fully connected layer is , To reduce the dimensionality ratio, the first fully connected layer uses the ReLU activation function. The weight of the second fully connected layer is , the second fully connected layer applies the Sigmoid activation function Channel attention weight Finally, the channel attention weight With feature map Multiply channel by channel to get the feature map after channel attention weighting .

[0090] The processing process of the spatial attention branch is: first, the feature map Perform maximum pooling and average pooling respectively to obtain and , and then concatenate the two according to the channel dimension to obtain . Then the convolution kernel is Convolutional layer and Sigmoid activation function , get the spatial attention weight Finally, the spatial attention weight With feature map Multiply each spatial position to obtain the feature map after spatial attention weighting .

[0091] Step S225: Input the feature map after channel attention weighting and the feature map after spatial attention weighting into the bidirectional heterogeneous feature interaction unit to obtain enhanced features corresponding to features of different scales.

[0092] Specifically, in the bidirectional heterogeneous feature interaction unit, the feature map after the channel attention weighting And the feature map after spatial attention weighting The fusion method is: first, the feature map after channel attention weighting And the feature map after spatial attention weighting Splice along the channel dimension, then apply a convolutional layer to the spliced ​​result to learn the interaction pattern of the two feature maps and obtain enhanced features . and They are two-dimensional convolution operations The convolution kernel size and the number of output channels.

[0093] Step S23: Input the multiple enhanced multi-scale features into a multi-scale feature fusion module to obtain fused features.

[0094] Specifically, the multi-scale feature fusion module is used to combine multiple enhanced features obtained by the deformable convolutional coding module Perform gradual fusion from top to bottom to obtain fusion features .

[0095] The specific fusion method is: first fusion from top to bottom, each enhanced feature Each corresponds to an intermediate fusion feature , for the lowest resolution enhanced feature ( ),have ; For other resolutions ( ), intermediate fusion features ,in Indicates the fusion feature of the previous lower resolution After upsampling, the enhanced features The splicing result of splicing, Indicates that the convolution operation is performed on the splicing result along the channel direction, and the number of output channels is .

[0096] In this embodiment, ， ， .

[0097] Finally, the intermediate fusion features are aggregated in the order of resolution from low to high through top-down upsampling, splicing and convolution to finally obtain the fusion features. In this embodiment as follows.

[0098] .

[0099] Step S24: Input the fused features into the mask generation module to obtain a predicted mask.

[0100] The beneficial effects of the target segmentation method proposed in this application through boundary learning optimization are mainly manifested in: 1. This application designs a loss function based on the Fourier descriptor during training. The Fourier descriptor can obtain a set of coefficients by performing a Fourier transform on the complex representation of the contour. These coefficients can describe the main features of the shape and are invariant to translation, rotation, and scaling. Therefore, even if the target moves in position, changes angle, or scales in size in the image, the Fourier descriptor can still remain consistent, making the model more robust when dealing with targets of different postures. Secondly, the Fourier descriptor can effectively capture the topological structure of the shape, not just the differences at the pixel level. By incorporating the Fourier descriptor loss into the loss function, the model is prompted to output a prediction mask that is not only close to the true value at the pixel level, but also more similar in shape features, thereby improving the prediction effect on small targets.

[0101] 2. This application further improves the adaptability to scale changes through normalization operation in the loss function based on Fourier descriptor.

[0102] 3. The low-frequency part of the Fourier descriptor mainly describes the overall structure of the shape, while the high-frequency part describes the details. This application retains the first K Fourier descriptors, which enables the model to focus on the main features of the shape and ignore some unnecessary details and noise.

[0103] 4. To better regulate the training process, this application sets adaptive weight parameters to balance the Fourier descriptor loss and the traditional MSE loss. By dynamically adjusting the weight values, the model can flexibly balance the importance of the two losses during training. Ultimately, the designed loss function can effectively improve the model's performance on shape features, especially when dealing with small object segmentation tasks. It can better capture the shape and topology of the target, not only enhancing the model's robustness and generalization ability, but also improving training efficiency and the quality of segmentation results.

[0104] 5. The object segmentation model in this application uses multi-scale feature extraction to obtain feature maps at multiple scales, which can simultaneously capture low-level local details and high-level global semantic information. Ultimately, it fully learns object information at different scales, thus better tackling object segmentation tasks in complex backgrounds.

[0105] 6. This application designs a deformable convolutional coding module combined with a dynamic sparse network: on the one hand, it can prompt the convolution kernel to dynamically adjust the sampling position according to the input features, so as to better adapt to the shape changes of the target in different rotation angles and different shooting angles, and more accurately capture the edges and details of these small targets to avoid omissions or mis-segmentation; on the other hand, the deformable convolutional coding module can capture global context information through the self-attention mechanism, so that the model can not only pay attention to local features, but also understand the overall structure and contextual relationship of the target, further enhancing the ability of feature representation; thirdly, by designing sparse attention selection, the number of sampling points of the deformable convolution can be reduced; at the same time, the combination of sparse deformable convolution and learnable gating unit can strengthen the features of key channels and spatial positions, improve the accuracy of model segmentation, and reduce the computational complexity of invalid features.

[0106] 7. This application uses FPN (Feature Pyramid Network) to achieve multi-scale feature fusion, transferring high-level semantic information to low-level feature maps while retaining more local detail information.

[0107] 8. The end-to-end trainable architecture used in this application can effectively capture small objects of different scales while enhancing robustness and generalization capabilities, ultimately adapting to the task of segmenting small objects in complex backgrounds and generating high-quality segmentation masks.

[0108] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0109] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0110] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0111] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A target segmentation method based on boundary learning optimization, characterized in that: The target segmentation method of the boundary learning optimization includes: Constructing and training a target segmentation model; the target segmentation model uses a loss function based on a pixel-level loss function, a Fourier descriptor loss function, and an adaptive weighting coefficient; the target segmentation model includes: a multi-scale feature extraction module, a deformable convolutional coding module, a multi-scale feature fusion module, and a mask generation module connected in sequence; Input the airport airspace image into the trained target segmentation model to obtain a prediction mask to complete the segmentation of the targets in the airport airspace; the prediction mask is the pixel position of the aircraft in the airport airspace image; The airport airspace image is input into the trained target segmentation model to obtain the predicted mask, which includes: Inputting the airport airspace image into the multi-scale feature extraction module to obtain features of different scales; Inputting features of different scales into the deformable convolutional coding module to obtain multiple enhanced features; Inputting the plurality of enhanced feature multi-scale features into the multi-scale feature fusion module to obtain fused features; The fused features are input into the mask generation module to obtain a predicted mask.

2. The target segmentation method of boundary learning optimization according to claim 1, characterized in that The deformable convolutional coding module includes: a channel-space separation offset prediction unit, a sparse attention selection unit, a deformable convolution unit, a learnable gating unit and a bidirectional heterogeneous feature interaction unit; Input features of different scales into the deformable convolutional coding module to obtain multiple enhanced features, including: Inputting features of different scales into the channel-space separation offset prediction unit to obtain spatial offsets and channel modulation weights corresponding to features of different scales; Inputting the channel modulation weight into the sparse attention selection unit for weight selection to obtain the sparse attention weight; Inputting the scale feature, the spatial offset, and the sparsified attention weight into the deformable convolution unit for convolution to obtain a feature map; Inputting the feature map into the learnable gating unit to obtain a channel attention weighted feature map and a spatial attention weighted feature map; The feature map weighted by channel attention and the feature map weighted by spatial attention are input into the bidirectional heterogeneous feature interaction unit to obtain enhanced features corresponding to features of different scales.

3. The target segmentation method of boundary learning optimization according to claim 2, characterized in that Based on the channel modulation weight, the channel modulation weight is input into the sparse attention selection unit for weight selection to obtain the sparse attention weight, specifically including: Calculating a channel modulation weight score based on the channel modulation weight; Selecting the channel modulation weight score to determine an index; constructing a sparse selection matrix based on the index; Based on the sparse selection matrix and the channel modulation weight, a sparse attention weight is obtained.

4. The target segmentation method of boundary learning optimization according to claim 3, characterized in that The calculation formula of the channel modulation weight score is: ； in, Score the channel modulation weight; is the channel modulation weight; For noise; is the temperature parameter.

5. The target segmentation method of boundary learning optimization according to claim 3, characterized in that The expression of the index is: ； in, is the index; For selection operation; Score the channel modulation weight; is the number of importance weights chosen.

6. The target segmentation method of boundary learning optimization according to claim 3, characterized in that The expression of the attention weight after sparsification is: ； in, is the attention weight after sparsification; Select the matrix for sparseness; is the dot product; is the channel modulation weight.

7. The target segmentation method of boundary learning optimization according to claim 2, characterized in that The expression of the feature map is: ； in, is the feature map; is the convolution operation of the deformable convolution unit; For the scale characteristics; is a deformable convolution kernel; is the attention weight after sparsification; is the spatial offset; for The field of real numbers, is the height of the feature map; is the width of the feature map; is the number of output channels.

8. The target segmentation method of boundary learning optimization according to claim 2, characterized in that The learnable gating unit includes: a channel attention branch and a spatial attention branch.

9. The target segmentation method of boundary learning optimization according to claim 1, characterized in that The training process of the target segmentation model specifically includes: Constructing a sample data pair; the sample data pair includes: a sample airport airspace image and a sample true mask; Inputting the sample airport airspace image into the target segmentation model to obtain a sample prediction mask; Based on the sample prediction mask and the corresponding sample true mask, the parameters of the target segmentation model are iteratively optimized using the loss function until the loss function reaches a minimum value or the iterative optimization round reaches a maximum value. The iterative optimization is stopped to obtain a trained target segmentation model.

10. The target segmentation method of boundary learning optimization according to claim 9, characterized in that: The expression of the loss function is: ； ； ； ； in, is the loss function; is the adaptive weighting coefficient; is the pixel-level loss function; is the Fourier descriptor loss function; Predict the mask for the sample pixel values; is the true mask of the sample pixel values; is the total number of pixels; is the total number of Fourier descriptors; is the true mask of the sample Fourier descriptors; is the 0th Fourier descriptor of the true mask of the sample; The first prediction mask for the sample Fourier descriptors; The 0th Fourier descriptor of the sample prediction mask; To take the model; is the number of current iteration training; is the total number of training iterations.

Images