Typical target edge detection and semantic segmentation method based on multi-task learning

Abstract

The invention discloses a typical target edge detection and semantic segmentation method based on multi-task learning. The method comprises the following steps: constructing a multi-task learning network based on edge detection and semantic segmentation; based on the training sample, completing training of a multi-task learning network, and constructing to obtain an edge detection and semantic segmentation multi-task model; and taking the remote sensing image containing the typical target as the input of the edge detection and semantic segmentation multi-task model, and outputting a typical target edge detection and semantic segmentation result through the edge detection and semantic segmentation multi-task model. According to the method, the semantic segmentation task and the edge detection task are combined, and the target fuzzy boundary identified in the semantic segmentation task is corrected through the edge geometric constraint information provided by the edge detection task, so that the precision and integrity of the segmentation result are improved; the problems of rough typical target segmentation boundary, structural feature deficiency and the like in the existing remote sensing image processing technology are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing image processing, and particularly relates to a typical target edge detection and semantic segmentation method based on multi-task learning. BACKGROUND

[0002] In the field of remote sensing applications, the rapid and accurate identification and fine analysis of typical targets in high-resolution optical remote sensing images have important strategic significance. For example, the identification of typical targets plays a key role in airport traffic management, flight scheduling, emergency rescue, etc. At present, the detection and identification of typical targets based on remote sensing images mainly rely on target detection techniques, such as R-CNN, YOLO, etc. series algorithms, which realize the preliminary positioning and classification of targets through bounding boxes. However, these methods are difficult to obtain the accurate contour and structural features of the target, which limits their application in high-precision scenarios. Semantic segmentation technology can realize the fine contour extraction and structural analysis of the target by classifying the image at the pixel level, and is an important basis for realizing high-precision identification of typical targets. However, due to the small size, diverse shape, and dense distribution of typical targets in remote sensing images, and often disturbed by complex backgrounds (such as shadows, buildings, other vehicles), it still faces great challenges to realize high-precision pixel-level semantic segmentation.

[0003] Although deep learning-based semantic segmentation models, such as U-Net, Deeplab series, etc., have made significant progress in this field, these models are usually designed for single tasks and focus more on the classification accuracy within the region, while the geometric structure constraints of the object boundary are weak. This leads to the problem of blurred boundaries and intra-class inconsistency in the segmentation results, which is difficult to meet the needs of high-precision applications. To overcome the above limitations, domestic and foreign scholars have begun to explore multi-task learning strategies, aiming to improve their performance through task cooperation and information complementation. In the semantic segmentation of remote sensing images, the edge information of the target plays a key role, making the joint modeling of semantic segmentation and edge detection a research hotspot in multi-task learning. Among them, edge detection can be regarded as a special case of semantic segmentation, and the model is trained to distinguish edges from backgrounds. Such multi-task learning can fully utilize the complementarity of semantic segmentation and edge detection: semantic segmentation provides class information, and edge detection provides geometric constraints, thereby improving the accuracy of semantic segmentation. On the one hand, the class information of semantic segmentation helps to suppress the background noise in remote sensing images, making edge detection more focused on the target area, realizing "semantic guiding edge"; on the other hand, the geometric constraints of edge detection can correct the blurred boundaries of semantic segmentation, realizing "edge-constrained semantics".

[0004] However, semantic segmentation and edge detection are a typical example of conflicting task combinations. Conflicting tasks refer to two tasks that, while closely related, differ significantly in feature dependencies, learning objectives, and optimization directions. Joint learning of these tasks can easily lead to task conflicts, resulting in decreased model performance. For example, edge detection is essentially a low-order feature extraction task, relying primarily on shallow visual cues such as local pixel intensity changes and being sensitive to details; while semantic segmentation is a high-order semantic understanding task, requiring the understanding of deep abstract features such as global context, target shape, and semantic relationships. Furthermore, edge detection emphasizes pixel-level precise localization, while semantic segmentation pursues accurate category prediction. The two compete at the target edge: the former requires sharp, distinguishable edges, while the latter tends towards blurred transitions to conform to the category region.

[0005] Existing research focuses primarily on the complementarity between tasks, while lacking a systematic exploration of the mechanisms for resolving task conflicts. Therefore, there is an urgent need for a new method that can effectively balance "information complementarity" and "task conflict" to improve the performance and robustness of multi-task learning in semantic segmentation of remote sensing images. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a typical target edge detection and semantic segmentation method based on multi-task learning. It combines the semantic segmentation task with the edge detection task. By using the edge geometric constraint information provided by the edge detection task, the blurred boundary of the target identified in the semantic segmentation task is corrected, thereby improving the accuracy and completeness of the segmentation results. This solves the problems of rough segmentation boundary and missing structural features of typical targets in existing remote sensing image processing technology.

[0007] To address the aforementioned technical problems, this invention discloses a typical target edge detection and semantic segmentation method based on multi-task learning, comprising: Construct a multi-task learning network based on edge detection and semantic segmentation; Based on the training samples, the multi-task learning network is trained, and a multi-task model for edge detection and semantic segmentation is constructed. Remote sensing images containing typical targets are used as input to a multi-task model for edge detection and semantic segmentation. The model outputs the edge detection and semantic segmentation results for typical targets.

[0008] In the aforementioned typical target edge detection and semantic segmentation methods based on multi-task learning, the multi-task learning network based on edge detection and semantic segmentation includes: The encoder is used to encode the features of the input remote sensing image, and extracts low-order features e1, mid-order features e2, high-order features e3 and semantic features e4 in layers; The edge detection decoder processes low-order feature e1, mid-order feature e2, and high-order feature e3 to obtain edge detection feature F. b ; The semantic segmentation decoder processes semantic feature e4 to obtain semantic segmentation feature F. s ; The conflict mitigation module is used to adjust the conflict based on edge detection features F. b and semantic segmentation features F s Information interaction based on conflict detection gating mechanism constraints is carried out to obtain enhanced edge detection features F'. b and enhanced semantic segmentation features F' s ; The output module is used to determine the enhanced edge detection features F'. b and enhanced semantic segmentation features F' s Output edge detection results and semantic segmentation results The total loss is calculated based on the adaptive weighted composite loss function.

[0009] In the aforementioned typical target edge detection and semantic segmentation methods based on multi-task learning, the edge detection decoder processes low-order feature e1, mid-order feature e2, and high-order feature e3 to obtain edge detection feature F. b This includes fusing high-order feature e3 and mid-order feature e2 to obtain fused feature F. b2 ; Fuse features F b2 The edge detection feature F is obtained by fusing it with the low-order feature e1. b .

[0010] In the aforementioned typical target edge detection and semantic segmentation methods based on multi-task learning, the semantic segmentation decoder processes the semantic feature e4 to obtain the semantic segmentation feature F. s This includes: performing feature optimization processing on semantic feature e4 to obtain the optimized feature tensor F. s4 ; the feature tensor F s4 The semantic segmentation feature F is obtained by fusing it with the corresponding low-order features. s .

[0011] In the aforementioned typical target edge detection and semantic segmentation methods based on multi-task learning, the conflict mitigation module, based on edge detection features F... b and semantic segmentation features F s Information interaction based on conflict detection gating mechanism constraints is carried out to obtain enhanced edge detection features F'. b and enhanced semantic segmentation features F' s At that time, it includes: edge detection features F b and semantic segmentation features Fs Channel concatenation is performed; after convolution and sigmoid activation, a "spatial conflict mask" is generated; a reverse attention mechanism is used, i.e., subtracting the "spatial conflict mask" from 1 to obtain the conflict detection gating attention matrix A; in the process of edge detection features guiding semantic segmentation features, the edge detection feature F... b After convolution and activation functions, we obtain the attention matrix C for "edge detection-guided semantic segmentation". Multiplying matrix C with matrix A yields the enhanced edge detection feature F'. b In the process of semantic segmentation features guiding edge detection features, semantic segmentation feature F s After convolution and activation functions, an attention matrix B for "semantic segmentation-guided edge detection" is obtained. Multiplying matrix B with matrix A yields the enhanced semantic segmentation feature F'. s .

[0012] In the above-mentioned typical target edge detection and semantic segmentation methods based on multi-task learning, the output module includes: an edge detection output module, a semantic segmentation output module, and an adaptive weighted composite loss function; The edge detection output module is used to determine the edge detection features F'. b Output edge detection results ; The semantic segmentation output module is used to perform semantic segmentation based on the enhanced semantic segmentation features F'. s Output semantic segmentation results ; An adaptively weighted composite loss function is used to calculate the total loss. Collaborative optimization of edge detection and semantic segmentation tasks.

[0013] In the aforementioned typical target edge detection and semantic segmentation methods based on multi-task learning,

[0014]

[0015]

[0016]

[0017] in, This represents the edge detection loss. Represents semantic segmentation loss. This indicates a loss of task consistency. This represents the binary cross-entropy loss function. Represents the Dice loss function. Indicates mean square error; Indicates edge label, Indicates semantic tags, Represents the Sobel operator; The weights represent the semantic segmentation loss. The weights represent the edge detection loss. The weight representing the consistency loss.

[0018] In the above-mentioned typical target edge detection and semantic segmentation method based on multi-task learning, training samples are paired in groups; each group of training samples includes three types of data: multi-channel optical remote sensing images containing targets cropped to a fixed size, binary images of the corresponding size with labeled target edges, and binary images of the corresponding size with labeled target as a whole.

[0019] The present invention has the following advantages: This invention discloses a typical target edge detection and semantic segmentation method based on multi-task learning. It combines the semantic segmentation task with the edge detection task. By using the edge geometric constraint information provided by the edge detection task, the blurred boundary of the target identified in the semantic segmentation task is corrected, thereby improving the accuracy and completeness of the segmentation results. This solves the problems of rough boundary and missing structural features in the segmentation of typical targets in existing remote sensing image processing technologies. Attached Figure Description

[0020] Figure 1 This is a flowchart of a typical target edge detection and semantic segmentation method based on multi-task learning in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of a multi-task learning network based on edge detection and semantic segmentation in an embodiment of the present invention; Figure 3 This is a schematic diagram of network front-end task decoupling in an embodiment of the present invention; Figure 4 This is a schematic diagram of a collision detection gating system in an embodiment of the present invention; Figure 5 This is a schematic diagram of an adaptive weighted composite loss function in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments disclosed in the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] One of the core ideas of this invention is to effectively alleviate the performance degradation caused by "task conflicts" in multi-task learning. This invention proposes a typical target edge detection and semantic segmentation method based on multi-task learning, which can achieve task decoupling. Specifically, it includes the following three levels of optimization design: First, a hierarchical feature sharing mechanism is adopted at the network front end, allowing the edge detection network and the semantic segmentation network to share only relatively low-order features from the backbone network, thereby reducing the risk of feature-level conflicts while ensuring information sharing between tasks. Second, a conflict detection gating mechanism is introduced in the task interaction part of the network middle. This mechanism can identify and enhance features compatible with each other's tasks during feature fusion, while suppressing features that may cause conflicts, thus achieving collaborative optimization between tasks. Finally, in the network output part, a task consistency loss function is designed to constrain the consistency between the target boundary predicted by the semantic segmentation task and the prediction result of the edge detection task, further improving the collaborative effect between tasks. In addition, this invention also designs an adaptive weighted composite loss function, which dynamically adjusts the weights of the losses for each task to achieve collaborative optimization of the segmentation task and the edge detection task, thereby minimizing the impact of task conflicts on model performance. To further enhance the "information complementarity" between tasks, this invention introduces feature sharing and feature interaction functions into the network structure, enabling the two tasks to complement each other's information and achieve collaborative optimization at different levels, thereby improving the segmentation accuracy and robustness of the overall model.

[0023] Reference Figure 1 In this embodiment, the typical target edge detection and semantic segmentation method based on multi-task learning includes: S1, construct a multi-task learning network based on edge detection and semantic segmentation.

[0024] In this embodiment, the multi-task learning network based on edge detection and semantic segmentation, after training, can perform edge detection and semantic segmentation of typical targets in remote sensing images. Here, the remote sensing image refers to a multi-channel optical remote sensing image. The dimensions of the remote sensing image are (C, H, W), where H and W represent the image height and width, respectively, and C represents the number of image channels.

[0025] like Figure 2 As shown, the multi-task learning network based on edge detection and semantic segmentation mainly includes: an encoder, an edge detection decoder, a semantic segmentation decoder, a conflict adjustment module, and an output module.

[0026] encoder An encoder refers to a feature extraction backbone network that enables hierarchical encoding of an input image, with changes in feature tensors between layers. For example... Figure 3As shown, taking the ResNet-50 backbone network as an example, it outputs four feature layers: layer e1 outputs (256, H / 4, W / 4), layer e2 outputs (512, H / 8, W / 8), layer e3 outputs (1024, H / 16, W / 16), and layer e4 outputs (2048, H / 32, W / 32). Layer e1 outputs low-order features, layer e2 outputs mid-order features, layer e3 outputs high-order features, and layer e4 outputs semantic features. That is, the encoder is mainly used to encode features from the input remote sensing image, extracting low-order features e1, mid-order features e2, high-order features e3, and semantic features e4 layer by layer.

[0027] Edge detection decoder Edge detection decoders rely on low-order features and cannot share only the lowest-level low-order features, otherwise "information complementarity" between tasks cannot be achieved; at the same time, they cannot share all layer features, otherwise it will cause conflicts with semantic segmentation tasks. Therefore, selecting appropriate low-order features and effectively fusing them achieves a balance between maximizing "information complementarity" and minimizing "task conflicts." Preferably, in this invention, fusing three layers of relatively low-order features—low-order feature e1, mid-order feature e2, and high-order feature e3—achieves an effective balance. That is, the edge detection decoder is mainly used to process low-order feature e1, mid-order feature e2, and high-order feature e3 to obtain edge detection feature F. b Specifically: the high-order feature e3 and the mid-order feature e2 are fused to obtain the fused feature F. b2 , fusion feature F b2 The dimensions are (64, H / 8, W / 8); furthermore, the fusion feature F b2 The edge detection feature F is obtained by fusing it with the low-order feature e1. b The dimensions are (64, H / 4, W / 4).

[0028] Semantic segmentation decoder The semantic segmentation decoder processes semantic feature e4 to obtain semantic segmentation feature F. s Specifically, the input semantic feature e4 can be optimized using the feature optimization module, outputting the optimized feature tensor F. s4 The dimensions are (256, H / 32, W / 32); further, the feature tensor F... s4 The semantic segmentation feature F is obtained by fusing it with the corresponding low-order features. s The dimensions are (64, H / 4, W / 4).

[0029] The feature optimization module is a module in the semantic segmentation decoder, which usually adopts atrous spatial pyramid pooling (ASPP). This module can be used to further extract multi-scale information from the features extracted from the backbone network using atrous convolution techniques.

[0030] Conflict resolution module The conflict mitigation module suppresses features that may cause conflicts by setting a conflict detection gating mechanism. Specifically, for example... Figure 4 As shown, the conflict adjustment module is specifically used to: adjust the edge detection features F b and semantic segmentation features F s Channel concatenation is performed; after convolution and a sigmoid activation function, a "spatial conflict mask" is generated. A larger value of this "spatial conflict mask" indicates a greater conflict between edge detection and semantic segmentation tasks. To reduce the weight of conflicting regions, a reverse attention mechanism is used, i.e., subtracting the "spatial conflict mask" from 1 to obtain the conflict detection gating attention matrix A. When tasks interact, matrix A will be applied to the interaction between tasks to regulate the conflict: in the process of edge detection features guiding semantic segmentation features, the edge detection feature F... b After convolution and activation functions, we obtain the attention matrix C for "edge detection-guided semantic segmentation". Multiplying matrix C with matrix A yields the enhanced edge detection feature F'. b Multiplying matrix C by matrix A enhances the weights of spatial locations in matrix C that are compatible with the semantic segmentation task, while suppressing conflicting spatial location weights. In the process of semantic segmentation features guiding edge detection features, semantic segmentation feature F... s After convolution and activation functions, an attention matrix B for "semantic segmentation-guided edge detection" is obtained. Multiplying matrix B with matrix A yields the enhanced semantic segmentation feature F'. s Multiplying matrix B by matrix A enhances the weights of spatial locations in matrix B that are compatible with edge detection tasks, while suppressing conflicting spatial location weights. Matrix B and matrix C simultaneously consider attention weights in both channel and spatial dimensions, allowing the model to adjust the importance of features in both dimensions, thereby better guiding edge segmentation and semantic segmentation.

[0031] Output module The output module is mainly used to determine the enhanced edge detection features F' b and enhanced semantic segmentation features F' s Output edge detection results and semantic segmentation results The total loss is calculated based on an adaptive weighted composite loss function. The output module mainly includes: The edge detection output module is used to determine the edge detection features F'. b Output edge detection results The edge detection output module outputs the edge detection results. It is a single channel, with dimensions of (1, H, W).

[0032] The semantic segmentation output module is used to perform semantic segmentation based on the enhanced semantic segmentation features F'. s Output semantic segmentation results The semantic segmentation output module outputs the semantic segmentation results. It is a single channel, with dimensions of (1, H, W).

[0033] An adaptively weighted composite loss function is used to calculate the total loss. This involves collaboratively optimizing edge detection and semantic segmentation tasks. For example... Figure 5 As shown, total loss It consists of three parts: edge detection loss Semantic segmentation loss Loss of task consistency .

[0034] Edge detection loss A linear combination of binary cross-entropy (BCE) loss and Dice loss is used to balance pixel-level classification accuracy and region overlap.

[0035] in, Indicates edge label, This represents the binary cross-entropy loss function. This represents the Dice loss function.

[0036] Similarly, semantic segmentation loss It also employs a linear combination of the same binary cross-entropy loss and Dice loss to balance pixel-level classification accuracy with region overlap:

[0037] in, Represents semantic tags.

[0038] Task consistency loss This invention reduces potential conflicts between the semantic segmentation task and the edge detection task by constraining the boundary output by mean squared error (MSE). The core idea is to extract target boundary features from the semantic segmentation result and compare them with the edge detection prediction result. This invention uses the Sobel operator to extract the target boundary of semantic segmentation.

[0039] in, Represents the Sobel operator. This represents the mean square error.

[0040] For dynamic balancing multi-task optimization, the total loss The composite loss function is designed to be calculated using the following adaptive weighting:

[0041] in, The weights representing the semantic segmentation loss (which can be learned) The weights representing the edge detection loss (which can be learned) The weight representing the consistency loss (fixed).

[0042] The adaptive weighted composite loss function design enables the multi-task learning network to automatically adjust the contribution of each task to the loss during training. That is, when the loss of a certain task decreases slowly, its corresponding weight decays exponentially, thereby avoiding dominating the optimization direction.

[0043] S2, based on the training samples, completes the training of the multi-task learning network and constructs a multi-task model for edge detection and semantic segmentation.

[0044] In this embodiment, training samples are paired in groups; each group of training samples includes three types of data: a multi-channel optical remote sensing image containing the target, cropped to a fixed size (H, W); a binary image of the target's edges labeled at the corresponding size (H, W); and a binary image of the target as a whole labeled at the corresponding size (H, W). During training, a batch of training samples is read in each time, and the batch size can be customized according to the training environment. After multiple rounds of training, a multi-task model for edge detection and semantic segmentation is obtained. The number of training rounds can be customized (e.g., 100 rounds), or training can be automatically terminated according to a preset stopping condition (e.g., the total loss no longer decreases significantly).

[0045] S3 takes a remote sensing image containing typical targets as input to the multi-task model for edge detection and semantic segmentation, and outputs the semantic segmentation results of typical targets through the multi-task model for edge detection and semantic segmentation.

[0046] In this embodiment, the multi-task model for edge detection and semantic segmentation can simultaneously perform edge detection and semantic segmentation tasks, with the corresponding outputs being enhanced typical target edge detection results and typical target semantic segmentation results, both of which are binary images. Optimally, to save energy consumption during model inference, this multi-task model can disable the output of a certain task, retaining only the necessary task branches. For example, given a remote sensing image as input, it can output only typical target semantic segmentation results or only typical target edge detection results. It is important to note that this multi-task model can accept optical remote sensing images of any spatial size as input, but the number of channels must remain consistent with the training samples. That is, when the image size input to the model is (C, H, W), C needs to remain consistent with the input during model training, while H and W are unrestricted.

[0047] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0048] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A typical target edge detection and semantic segmentation method based on multi-task learning, characterized in that, include: Construct a multi-task learning network based on edge detection and semantic segmentation; Based on the training samples, the multi-task learning network is trained, and a multi-task model for edge detection and semantic segmentation is constructed. Remote sensing images containing typical targets are used as input to a multi-task model for edge detection and semantic segmentation. The model outputs the edge detection and semantic segmentation results for typical targets.

2. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 1, characterized in that, Multi-task learning networks based on edge detection and semantic segmentation include: The encoder is used to encode the features of the input remote sensing image, and extracts low-order features e1, mid-order features e2, high-order features e3 and semantic features e4 in layers; The edge detection decoder processes low-order feature e1, mid-order feature e2, and high-order feature e3 to obtain edge detection feature F. b ; The semantic segmentation decoder processes semantic feature e4 to obtain semantic segmentation feature F. s ; The conflict mitigation module is used to adjust the conflict based on edge detection features F. b and semantic segmentation features F s Information interaction based on conflict detection gating mechanism constraints is carried out to obtain enhanced edge detection features F'. b and enhanced semantic segmentation features F' s ; The output module is used to determine the enhanced edge detection features F'. b and enhanced semantic segmentation features F' s Output edge detection results and semantic segmentation results The total loss is calculated based on the adaptive weighted composite loss function.

3. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 2, characterized in that, The edge detection decoder processes low-order feature e1, mid-order feature e2, and high-order feature e3 to obtain edge detection feature F. b This includes fusing high-order feature e3 and mid-order feature e2 to obtain fused feature F. b2 ; Fuse features F b2 The edge detection feature F is obtained by fusing it with the low-order feature e1. b .

4. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 2, characterized in that, The semantic segmentation decoder processes semantic feature e4 to obtain semantic segmentation feature F. s This includes: performing feature optimization processing on semantic feature e4 to obtain the optimized feature tensor F. s4 ; the feature tensor F s4 The semantic segmentation feature F is obtained by fusing it with the corresponding low-order features. s .

5. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 2, characterized in that, The conflict mitigation module is based on edge detection features F b and semantic segmentation features F s Information interaction based on conflict detection gating mechanism constraints is carried out to obtain enhanced edge detection features F'. b and enhanced semantic segmentation features F' s At that time, it includes: edge detection features F b and semantic segmentation features F s Channel concatenation is performed; after convolution and sigmoid activation, a "spatial conflict mask" is generated; a reverse attention mechanism is used, i.e., subtracting the "spatial conflict mask" from 1, to obtain the conflict detection gating attention matrix A; in the process of edge detection features guiding semantic segmentation features, the edge detection feature F... b After convolution and activation functions, we obtain the attention matrix C for "edge detection-guided semantic segmentation". Multiplying matrix C with matrix A yields the enhanced edge detection feature F'. b In the process of semantic segmentation features guiding edge detection features, semantic segmentation feature F s After convolution and activation functions, an attention matrix B for "semantic segmentation-guided edge detection" is obtained. Multiplying matrix B with matrix A yields the enhanced semantic segmentation feature F'. s .

6. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 2, characterized in that, The output module includes: an edge detection output module, a semantic segmentation output module, and an adaptive weighted composite loss function; The edge detection output module is used to determine the edge detection features F'. b Output edge detection results ; The semantic segmentation output module is used to perform semantic segmentation based on the enhanced semantic segmentation features F'. s Output semantic segmentation results ; An adaptively weighted composite loss function is used to calculate the total loss. Collaborative optimization of edge detection and semantic segmentation tasks.

7. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 2, characterized in that, in, This represents the edge detection loss. Represents semantic segmentation loss. This indicates a loss of task consistency. This represents the binary cross-entropy loss function. Represents the Dice loss function. Indicates mean square error; Indicates edge label, Indicates semantic tags, Represents the Sobel operator; The weights represent the semantic segmentation loss. The weights represent the edge detection loss. The weight representing the consistency loss.

8. The typical target edge detection and semantic segmentation method based on multi-task learning according to claim 1, characterized in that, Training samples are paired in groups; each training sample group includes three types of data: a multi-channel optical remote sensing image containing the target cropped to a fixed size, a binary image of the target edge labeled at the corresponding size, and a binary image of the target as a whole labeled at the corresponding size.

Images