A water body boundary identification method and device based on boundary perception collaborative optimization

Abstract

The application provides a water body boundary identification method and device based on boundary perception collaborative optimization, belongs to the field of artificial intelligence remote sensing SAR image segmentation, and comprises the following steps: a deep learning model is constructed, a real-time double-branch semantic segmentation framework is adopted to meet the timeliness requirement of water body boundary identification; an auxiliary boundary prediction branch is introduced to highlight high-frequency semantic information, a boundary detection is taken as an optimization target, and a boundary perception loss is introduced to predict complex water body boundaries; a pixel attention module, a context fast aggregation module and a boundary attention guiding module are proposed to mine spatial detail information, context information and boundary information of a target image respectively, control effective learning of context semantic information, guarantee reliability and timeliness of extracted information, guide effective fusion of various information at a boundary area, jointly optimize original double-branch and auxiliary branches, and realize accurate identification of water body boundaries. The application can improve identification precision.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence remote sensing SAR (synthetic aperture radar) image segmentation application, specifically involving a water body boundary recognition method and device based on boundary perception collaborative optimization. Background Technology

[0002] Water body boundary identification is crucial for understanding the distribution and changes of water resources. This includes extracting parameters such as the location, area, shape, and width of water bodies like lakes, rivers, and reservoirs, which is essential for water resource surveys and macro-monitoring. Examples include flood boundary identification in emergency disaster relief, drought and water shortage assessment, and important applications in environmental protection, coastal zone monitoring, and mapmaking. A key research direction is how to accurately identify and automate water boundaries in complex environments using radar imaging, and how to output the latitude and longitude coordinates of boundaries in real time from radar images. Currently, scholars both domestically and internationally have conducted extensive research on water body boundary identification. In traditional methods, surface water surveys are significant for understanding coastline changes, environmental protection, disaster prevention and mitigation, and water quality monitoring. Remote sensing images can quickly, repeatedly, and accurately acquire the spatiotemporal distribution characteristics of surface water. Thresholding methods, which select thresholds for image segmentation by analyzing the spectral curves of water bodies and background features, are simple to operate but have a low signal-to-noise ratio and are prone to confusing water bodies with background features. While decision tree methods and automatic water body extraction methods have addressed the significant shortcomings of thresholding methods, they still face challenges in improving accuracy.

[0003] From a data acquisition perspective, deep learning in water boundary identification can be divided into single-source data and multi-source data applications. Single-source data refers to data acquired using a single payload (such as radar, visible light, or infrared imagery). Visible light and infrared imagery are difficult to apply in practical engineering because they cannot transmit the latitude and longitude coordinates of two-dimensional images; radar imagery, however, can obtain latitude and longitude information and is suitable for practical engineering. Multi-source data involves the fusion of multiple image types, but in practice, it is difficult to align the field of view and image center point of visible light, infrared, and radar imagery.

[0004] From a model training perspective, deep learning is divided into supervised and unsupervised training. While unsupervised training reduces reliance on labeled data, it places high demands on model performance. Currently, the multimodal field is developing rapidly. Therefore, water boundary recognition mainly uses single-source data for supervised semantic segmentation model training, outputting pixel-by-pixel boundary segmentation results. However, the high-density output leads to a decrease in recognition speed. Summary of the Invention

[0005] To address the problems of water bodies being easily confused with the background, difficulty in distinguishing complex backgrounds, poor timeliness, and poor algorithm robustness, this invention provides a water body boundary recognition method and device based on boundary-aware collaborative optimization. It proposes a deep learning model-based water body boundary recognition algorithm based on boundary collaborative optimization, designs a real-time dual-branch semantic segmentation framework, and constructs an auxiliary boundary prediction branch on this basis. High-frequency semantic information is emphasized, and boundary detection is used as the branch objective. This achieves joint optimization with the spatial detail extraction branch and the context information extraction branch, outputting more accurate boundary recognition results.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A water body boundary identification method based on boundary-aware collaborative optimization includes:

[0008] A deep learning model is constructed, employing a real-time dual-branch semantic segmentation framework. The deep learning model includes a spatial detail information extraction branch and a context information extraction branch, which are used to meet the timeliness requirements of water body boundary identification.

[0009] An auxiliary boundary prediction branch is introduced to highlight high-frequency semantic information. Boundary detection is used as the optimization objective, and a boundary-aware loss is introduced to predict complex water body boundaries.

[0010] A pixel attention module, a context fast aggregation module, and a boundary attention guidance module are constructed to mine spatial detail information, context information, and boundary information of the target image, respectively. The effective learning of context semantic information is controlled to ensure the reliability and timeliness of the extracted information and guide the effective fusion of various information at the boundary region. The real-time bi-branch semantic segmentation framework and the auxiliary boundary prediction branch are jointly optimized to achieve accurate identification of water body boundaries.

[0011] The present invention also provides a water body boundary identification device based on boundary perception and collaborative optimization, comprising the following modules:

[0012] The model building module adopts a real-time dual-branch semantic segmentation framework. The deep learning model includes a spatial detail information extraction branch and a context information extraction branch to meet the timeliness requirements of water body boundary identification.

[0013] The prediction module introduces an auxiliary boundary prediction branch to highlight high-frequency semantic information, with boundary detection as the optimization objective, and introduces boundary-aware loss to predict complex water body boundaries.

[0014] The recognition module comprises a pixel attention module, a context fast aggregation module, and a boundary attention guidance module, which are used to mine spatial detail information, context information, and boundary information of the target image, respectively. It controls the effective learning of context semantic information, realizes the reliability and timeliness of the extracted information, and guides the effective fusion of various information at the boundary region. It jointly optimizes the real-time bi-branch semantic segmentation framework and the auxiliary boundary prediction branch to achieve accurate recognition of water body boundaries.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for water boundary recognition based on boundary-aware collaborative optimization.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for water body boundary identification based on boundary-aware collaborative optimization.

[0017] Beneficial effects:

[0018] 1. In this invention, complex boundary segmentation is more accurate. The model of this invention constructs boundary branches and optimizes the loss, which can better handle the problem of background and boundary being easily confused.

[0019] 2. The model structure of this invention is small and relatively fast. When processing 1024×1024 data, the model's processing time on a graphics card with a computing power of 35.6 TFLOPS@FP32 is only 8ms.

[0020] 3. This invention provides a complete deployment scheme and introduces the image coordinate and latitude / longitude alignment method in practical engineering applications, which can more effectively carry out engineering deployment. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of a deep learning model according to an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram of a pixel attention module;

[0023] Figure 3 This is a schematic diagram of the parallel structure of the context fast aggregation module;

[0024] Figure 4 A schematic diagram of the boundary attention guidance module;

[0025] Figure 5 A flowchart for data testing;

[0026] Figure 6 The result image obtained for inference by the deep learning model;

[0027] Figure 7 This is a flowchart of a water body boundary identification method based on boundary-aware collaborative optimization, according to an embodiment of the present invention.

[0028] Figure 8 This is a schematic diagram of a water boundary recognition device based on boundary perception and collaborative optimization, according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0030] In water body boundary monitoring scenarios, accurate and timely feedback of the boundaries and area of ​​water bodies formed by disasters is required. However, these boundaries are often complex. Therefore, such as Figure 7 As shown, an embodiment of the present invention provides a water boundary identification method based on boundary-aware collaborative optimization, comprising:

[0031] Step 1: Considering the timeliness of water body boundary identification, a deep learning model is constructed, and a real-time dual-branch semantic segmentation framework is adopted, including a spatial detail information extraction branch (i.e., the spatial detail branch below) and a context information extraction branch (i.e., the low-frequency extraction branch below).

[0032] Step 2: In order to predict complex water body boundaries, an auxiliary boundary prediction branch (hereinafter referred to as the auxiliary boundary prediction branch) is proposed, which highlights high-frequency semantic information, takes boundary detection as the optimization objective, and introduces boundary-aware loss.

[0033] The semantics of pixels within a segmented target are consistent, and semantic inconsistencies only occur at the boundaries of adjacent objects. Therefore, the goal of the boundary prediction branch is boundary detection.

[0034] Step 3: In order to fully exploit the spatial details, contextual information and boundary information of the target image, the original dual-branch and auxiliary branches are jointly optimized. A pixel attention module, a context fast aggregation module and a boundary attention guidance module are proposed respectively. This fully exploits the spatial details, contextual information and boundary information of the target image, and jointly optimizes the original dual-branch and auxiliary branches to achieve accurate identification of water body boundaries.

[0035] The pixel attention module controls the effective learning of contextual semantic information and ensures that the original information is not overwhelmed. The context fast aggregation module ensures the reliability and timeliness of the extracted information. The boundary attention guidance module guides the effective fusion of various information at the boundary region, ensuring semantic accuracy while preserving spatial and geometric details. These modules improve the accuracy of water body boundary recognition.

[0036] Specifically, step 1 includes:

[0037] The structure of the deep learning model proposed in this invention is as follows: Figure 1 As shown, in order to balance the accuracy and timeliness of the intelligent recognition module, a real-time dual-branch semantic segmentation framework is adopted, with high-resolution SAR images as input.

[0038] The deep learning model is:

[0039] The input serves as the starting point for the deep learning model. It passes through two convolutional layers, each using a 3×3 kernel with the same number of channels. Then it goes through two convolutional blocks, each containing one channel. Next, it passes through two more convolutional blocks, each containing two channels. Finally, it proceeds to the spatial detail branch, the low-frequency extraction branch, and the auxiliary boundary prediction branch.

[0040] In the spatial detail branch, the data passes through two convolutional blocks, each containing two channels, and then through an attention block (Pag) to apply an attention mechanism. Data from the low-frequency extraction branch passes through a 1×1 convolutional kernel, three convolutional blocks, and a 3×3 convolutional kernel. The data passes through a 1×1 convolutional kernel with 4 input channels and 2 output channels, and then together with the data that passed through the attention block, passes through two more convolutional blocks, each containing two channels. It then passes through the attention block again, applying the attention mechanism. Finally, together with the data that passed through three convolutional blocks, each containing eight channels, and then through a 1×1 convolutional kernel with 8 input channels and 2 output channels, it passes through a single convolutional block containing two channels.

[0041] In the low-frequency extraction branch, the data passes through three convolutional blocks, each containing four channels, and then through three convolutional blocks, each containing eight channels, with a 3×3 convolutional kernel, each containing eight channels, and then through three convolutional blocks, each containing eight channels, with a 3×3 convolutional kernel, each containing eight channels, and then through three convolutional blocks, each containing eight channels, and then through three convolutional blocks, each containing eight channels, with a 3×3 convolutional kernel, each containing eight channels, and then through three convolutional blocks, each containing eight channels, with a 3×3 convolutional kernel, each containing four channels, and 128 channels, and finally through a 1×1 convolutional kernel, with 128 channels and the number of channels equal to the number of categories, to achieve the final output.

[0042] Specifically, in step 2, in the auxiliary boundary prediction branch, the data passes through a convolutional block containing two channels, the output of the low-frequency extraction branch is added to the output of the auxiliary boundary prediction branch, and then passes through a convolutional block containing two channels, the output of the low-frequency extraction branch is added to the output of the auxiliary boundary prediction branch again, and then passes through a convolutional block containing two channels.

[0043] The outputs from the spatial detail branch, the low-frequency extraction branch, and the auxiliary boundary prediction branch all enter the boundary attention guidance module (Bag) (Bag is...) Figure 1 The attention block below the pyramid pooling in the middle, Figure 1 The other attention blocks are all Pag), then it enters a 3×3 convolution kernel with 4 input channels and 128 output channels, then enters a 1×1 convolution kernel with 128 input channels and the number of output channels equal to the number of categories, and finally outputs the result.

[0044] Specifically, step 3 includes:

[0045] First, to ensure that the DB branch (spatial detail information extraction branch) selectively learns useful semantic features from the low-frequency CB branch (low-frequency context information extraction branch) without being overwhelmed, a pixel attention module (Pag) is proposed, with the specific structure as follows: Figure 2As shown. The rich and accurate semantic information provided by the spatial detail branch is crucial for detail parsing and boundary detection in the low-frequency extraction branch and the ABB branch (Auxiliary boundary prediction branch). Both the low-frequency extraction branch and the ABB branch contain relatively few layers and channels, so the CB branch is considered a backup to the other two branches.

[0046] To ensure that the DB branch selectively learns useful semantic features from the CB branch without being overwhelmed, a pixel attention module (Pag) is proposed with the following specific structure: Figure 2 As shown. The rich and accurate semantic information provided by the CB branch is crucial for detailed parsing and boundary detection of the DB and ABB branches. Both the DB and ABB branches contain relatively few layers and channels, so the CB branch can be considered a backup of the other two branches.

[0047] Figure 2 In this process, the input image is divided into two branches for processing: the spatial detail branch and the low-frequency extraction branch.

[0048] The input image is first processed through a spatial detail branch, where spatial detail features are extracted using convolution operations. These feature maps are then downsampled to 1 / 8 of their original size. Simultaneously, the input image is also processed through a low-frequency extraction branch, where low-frequency features are extracted using linear correction units and convolutional modules. These feature maps are also downsampled to 1 / 8 of their original size.

[0049] The feature maps from both branches are fed into the pixel attention module. In this module, the feature maps are first passed through a convolutional layer, then through a sigmoid activation function to generate attention weights. These attention weights are used to weight the feature maps, highlighting important features and suppressing unimportant ones. The two weighted feature maps output from the pixel attention module are then multiplied element-wise (…). The feature map is then combined with the original feature map. The combined feature map is fused using an element-wise addition (⊕) operation. The fused feature map is then upsampled to its original size. The upsampled feature map is further processed by a convolutional module to integrate and refine the features. Finally, the feature map processed by the convolutional module is used as the output for subsequent image analysis or processing tasks.

[0050] This invention enhances the representation of important features by combining spatial details and low-frequency features, utilizing an attention mechanism, and integrating and refining features through upsampling and convolution operations, thereby improving the accuracy and efficiency of image processing.

[0051] Define the vectors of corresponding pixels in the feature maps of the DB and CB branches as follows: and ,but The output of the activation function can be expressed as:

[0052] ;

[0053] in, This indicates the probability that the two pixels belong to the same object. and For 1×1 convolutions with different numbers of channels. If Higher levels of trust will lead to greater trust. Because the CB branch is semantically rich and accurate, it is more likely to be trusted if the value is low. Therefore, the output of the pixel attention module It can be written as:

[0054] ;

[0055] Secondly, to construct prior information for the global scene, a fast context aggregation module (PAPPM) is proposed. Traditional deep aggregation pyramid pooling modules (DAPPM) cannot be parallelized in their computation, which is very time-consuming. Furthermore, each scale in the pyramid pooling module contains too many channels, potentially exceeding the representational capacity of lightweight models. Therefore, PAPPM improves upon this framework by modifying its connection method to enable parallelization, such as... Figure 3 As shown, the number of channels for each scale was reduced from 128 to 96. Figure 3In the diagram, (5,2)-average pooling represents average pooling with a kernel size of 5×5, a stride of 2, and a patch size of 2. Specifically, a 1×1 convolution operation is performed on the input data to adjust the number of channels or perform feature transformation. (5,2)-average pooling represents a 5×5 region average pooling operation on the input data with a stride of 2. It is a downsampling operation used to reduce the spatial dimension of the data. Convolution [1×1], upsampling represents performing a 1×1 convolution on the average pooled data and then upsampling to restore the original spatial dimension. (9,4)-average pooling represents a 9×9 region average pooling operation on the input data with a stride of 4. This is another downsampling operation. Convolution [1×1], upsampling represents performing a 1×1 convolution on the average pooled data and then upsampling. (17,8)-average pooling represents a 17×17 region average pooling operation on the input data with a stride of 8. Convolution [1×1], upsampling means performing a 1×1 convolution on the average pooled data, followed by upsampling. Global average pooling means performing global average pooling on the input data, compressing the feature map of each channel into a single value. Then, convolution [1×1], upsampling means performing a 1×1 convolution on the globally average pooled data, followed by upsampling. After summation, convolution [3×3] is performed separately, meaning a 3×3 convolution operation is performed on the data of each branch, which is usually used to capture local features. Then, concatenation by channel means concatenating the outputs of all branches along the channel dimension, followed by a 1×1 convolution operation to integrate features. A final 1×1 convolution operation is performed before the output to adjust the number of output channels or to perform a final non-linear transformation. Finally, the processed data is output.

[0056] This invention utilizes a multi-branch convolutional neural network module, employing average pooling and upsampling operations at different scales, combined with 1×1 and 3×3 convolutions, to extract and fuse multi-scale features. Finally, these features are integrated through concatenation and convolution operations to output the final result. This structure helps capture multi-scale information in images and improves the model's expressiveness.

[0057] Finally, a boundary attention guidance module (Bag) is proposed. Given the boundary features extracted by the ABB branch, boundary attention is used to guide the fusion of spatial detail features and contextual features to achieve better semantic segmentation results. Although the contextual branch has semantic accuracy, it loses too much spatial and geometric detail in the boundary region. Therefore, the detail branch is used to provide better spatial detail, and the model is forced to trust the detail branch more in the boundary region, while contextual features are used to fill other regions. As shown in Figure 4, high-frequency and low-frequency regions are filled with detail features and contextual features, respectively.

[0058] Figure 4In this process, the input image first undergoes spatial detail extraction via a spatial detail branch. This branch contains convolutional operations to capture local features in the image. Gray markers indicate semantic errors. The low-frequency extraction branch is responsible for extracting low-frequency information from the image, using different filters or convolutional kernels. Figure 4 The two-dimensional matrix in the image represents the processed feature map. The auxiliary boundary prediction branch is used to predict boundary or edge information in the image. Figure 4 A matrix with probability values ​​(0.5) is shown, representing the confidence or probability of the boundary prediction. The feature maps extracted from the three branches are processed by addition (⊕) and multiplication (…). The feature maps are then fused using a 3×3 convolution operation. The final feature map is further processed through this convolution. Convolution is used to refine the feature map, extract higher-level features, or perform feature dimensionality reduction. After these processes, the final output feature map is obtained, which can be used for subsequent image segmentation, classification, or other computer vision tasks.

[0059] Let the vectors of the corresponding pixels in the feature maps of the DB branch, CB branch, and ABB branch be respectively... , and ,but Output and the output of the boundary attention guidance module It can be represented as:

[0060] ;

[0061] ;

[0062] in, It is a combination of convolution, batch normalization, and ReLU (linear correction module); This represents the multiplication of two convolutional feature vectors.

[0063] The loss function of a deep learning model is a composite function consisting of four parts, including: adding a semantic head at the output of the first boundary attention-guided module to generate additional semantic loss. To better optimize the entire network; to handle the imbalance problem in boundary detection, a weighted binary cross-entropy loss is used. ; and Represents the cross-entropy loss (CE), and simultaneously affects... We utilize boundary-aware CE loss, leverage the output of the boundary head to coordinate semantic segmentation and boundary detection tasks, and enhance the functionality of the boundary attention guidance module.

[0064] ;

[0065] in, It's a real label. It is the probability predicted by the model. When the true label... When the probability is 1, if the predicted probability is... If the value is close to 1, the loss is close to 0; conversely, if... When the value approaches zero, the loss becomes extremely large. Similarly, when... When it is 0, if If it approaches 0, then the loss approaches 0; if If the value is close to 1, the loss will become very large.

[0066] ;

[0067] in, It's a real label. It is the probability predicted by the model.

[0068] ;

[0069] in, This represents a predefined threshold. , and These are the segmentation classes. The boundary header output of each pixel, the segmentation of the ground truth value, and the prediction result; i represents the value of the i-th pixel, and c represents the c-th channel.

[0070] ;

[0071] in, , , , For each pixel value, a different constant value is assigned based on the contribution of each loss.

[0072] By designing three core modules—pixel attention module, context fast aggregation module, and boundary attention guidance module—and a composite loss function, the boundary detection capability of the real-time dual-branch semantic segmentation framework can be extended. This enables spatial detail information, context information, and boundary information to guide each other and jointly optimize within the network, thereby achieving the identification of complex water body boundaries.

[0073] The data used to train the deep learning model is labeled, with water areas identified by aligning the water bodies using a map while labeling. Due to the large image size, image tiling is required for training, with each tiling image measuring 1024×1024 pixels. The images do not need to overlap during tiling, and the final output is a binary image for semantic segmentation.

[0074] During testing, a real-world application environment is simulated, such as... Figure 5 As shown, radar images are first acquired. The images acquired by the actual radar imaging payload are in stripe mode, with a size ranging from 10000×10000. Directly reducing the image size would result in significant loss of detail. Therefore, during testing, the image is cropped to a size of 1024×1024, consistent with the input size of the deep learning model trained above. Testing is then performed, and a binary image is output. The output binary images are then stitched together to recreate the original size (i.e.,...). Figure 5 The composite large image is obtained by extracting the latitude and longitude information of the image and aligning the contour pixels (i.e., ...). Figure 5 The algorithm finds the coordinates of all closed regions and outputs the results. gdal is a Python library specifically designed to read the latitude and longitude information of radar images (i.e., coordinates of all closed regions). Figure 5 (Mapping latitude and longitude).

[0075] like Figure 6 The image shown is a binary image; the river region is displayed in white, and all other regions are displayed in black, allowing the river portion to be extracted.

[0076] The test times of the constructed deep learning model on a graphics card with a computing power of 35.6 TFLOPS@FP32 are shown in Table 1. Post-processing takes up a long time, including finding image contours and latitude and longitude matching. Latitude and longitude matching is performed pixel by pixel, which is relatively time-consuming.

[0077] Table 1

[0078]

[0079] The latitude and longitude information is output pixel by pixel, eliminating the need for gradual pixel alignment. The latitude and longitude are output every 10 pixels, saving approximately 10ms of time.

[0080] This invention uses an imaging radar payload as an image acquisition tool. A manned aircraft carrying the radar payload conducts aerial photography of some waters in China. The data area range is shown in Table 2. The entire radar image area map is acquired and annotated. Based on the requirements of accuracy and real-time performance, a deep learning model is constructed. An auxiliary boundary prediction branch is proposed on the basis of a real-time dual-branch semantic segmentation framework to highlight high-frequency semantic information. With boundary detection as the branch objective, joint optimization is completed with the spatial detail extraction branch and the context information extraction branch to output more accurate boundary recognition results.

[0081] Table 2

[0082]

[0083] like Figure 8As shown, the present invention also provides a water body boundary identification device based on boundary perception and collaborative optimization, comprising the following modules:

[0084] The model building module constructs a deep learning model and adopts a real-time dual-branch semantic segmentation framework, including a spatial detail information extraction branch and a context information extraction branch, to meet the timeliness requirements of water body boundary identification.

[0085] The prediction module introduces an auxiliary boundary prediction branch, which highlights high-frequency semantic information, takes boundary detection as the optimization objective, and introduces boundary-aware loss to predict complex water body boundaries.

[0086] The recognition module proposes a pixel attention module, a context fast aggregation module, and a boundary attention guidance module, which are used to mine spatial detail information, context information, and boundary information of the target image, respectively. They control the effective learning of context semantic information, ensure the reliability and timeliness of the extracted information, and guide the effective fusion of various information at the boundary region. They jointly optimize the original dual branches and auxiliary branches to achieve accurate recognition of water body boundaries.

[0087] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for water boundary recognition based on boundary-aware collaborative optimization.

[0088] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for water body boundary identification based on boundary-aware collaborative optimization.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A water body boundary identification method based on boundary-aware collaborative optimization, characterized in that, include: A deep learning model is constructed, employing a real-time dual-branch semantic segmentation framework. The deep learning model includes a spatial detail information extraction branch and a context information extraction branch, which are used to meet the timeliness requirements of water body boundary identification. An auxiliary boundary prediction branch is introduced to highlight high-frequency semantic information. Boundary detection is used as the optimization objective, and a boundary-aware loss is introduced to predict complex water body boundaries. A pixel attention module, a context fast aggregation module, and a boundary attention guidance module are constructed to mine spatial detail information, context information, and boundary information of the target image, respectively. The effective learning of context semantic information is controlled to ensure the reliability and timeliness of the extracted information and guide the effective fusion of various information at the boundary region. The real-time bi-branch semantic segmentation framework and the auxiliary boundary prediction branch are jointly optimized to achieve accurate identification of water body boundaries.

2. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 1, characterized in that, The loss function of the deep learning model includes semantic loss, weighted binary cross-entropy loss, and boundary-aware cross-entropy loss. By setting the weights of semantic loss, weighted binary cross-entropy loss, and boundary-aware cross-entropy loss, the semantic segmentation and boundary detection tasks are coordinated.

3. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 1, characterized in that, The data used to train the deep learning model was labeled by aligning the water areas with a map while simultaneously labeling the data.

4. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 1, characterized in that, The goal of the boundary prediction branch is boundary detection.

5. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 1, characterized in that, The rich and accurate semantic information provided by the context information extraction branch is used for detail parsing and boundary detection of the spatial detail information extraction branch and the auxiliary boundary prediction branch; the context information extraction branch is regarded as a backup of the spatial detail information extraction branch and the auxiliary boundary prediction branch.

6. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 5, characterized in that, Given the boundary features extracted by the auxiliary boundary prediction branch, boundary attention is used to guide the fusion of spatial detail features and contextual features.

7. The water body boundary identification method based on boundary-aware collaborative optimization according to claim 1, characterized in that, The pixel attention module is used to control the effective learning of contextual semantic information and ensure that the original information is not overwhelmed; the context fast aggregation module is used to ensure the reliability and timeliness of the extracted information; the boundary attention guidance module is used to guide the effective fusion of various information at the boundary region, ensuring the semantic accuracy of the boundary region while preserving spatial and geometric details.

8. A water body boundary recognition device based on boundary perception and collaborative optimization, characterized in that, Includes the following modules: The model building module adopts a real-time dual-branch semantic segmentation framework. The deep learning model includes a spatial detail information extraction branch and a context information extraction branch to meet the timeliness requirements of water body boundary identification. The prediction module introduces an auxiliary boundary prediction branch to highlight high-frequency semantic information, with boundary detection as the optimization objective, and introduces boundary-aware loss to predict complex water body boundaries. The recognition module comprises a pixel attention module, a context fast aggregation module, and a boundary attention guidance module, which are used to mine spatial detail information, context information, and boundary information of the target image, respectively. It controls the effective learning of context semantic information, realizes the reliability and timeliness of the extracted information, and guides the effective fusion of various information at the boundary region. It jointly optimizes the real-time bi-branch semantic segmentation framework and the auxiliary boundary prediction branch to achieve accurate recognition of water body boundaries.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the water body boundary recognition method based on boundary-aware collaborative optimization as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of a water boundary identification method based on boundary-aware collaborative optimization as described in any one of claims 1-7.

Images