A neural network architecture for boundary and morphology characterization

Abstract

The application discloses a neural network architecture for boundary and morphology representation, and relates to the technical field of deep learning. A boundary representation branch encodes input feature tensors in multi-scale and multi-directional responses based on multiple directional geometry modules, enhances the perception of multi-scale and multi-directional boundary information, and improves the continuity and stability of boundary representation. A shape semantic modeling branch performs hierarchical context modeling and structure representation enhancement based on multiple convolution residual units and multiple morphology space perception adaptive units, improving the representation ability of the overall shape and context relationship of the region. A feature fusion branch further performs cross-layer interaction, weight modulation and double-flow combination processing on the boundary representation tensor and the region structure representation tensor, strengthening the interaction and fusion between different hierarchical features. Thus, the problems of boundary fracture, local false response, structure discontinuity and insufficient morphology regularity in the prior art can be effectively alleviated, and the output structure integrity, boundary stability and representation consistency can be improved.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, specifically to a neural network architecture oriented towards boundary and morphological representation. Background Technology

[0002] Farmland is a crucial spatial unit in agricultural production, resource surveys, and arable land monitoring. Its boundary location, morphological integrity, and structural regularity directly impact the accuracy of arable land area calculations, precision agriculture input allocation, land consolidation assessments, and arable land protection monitoring. With the continuous improvement of high-resolution remote sensing data acquisition capabilities, the automatic representation of farmland boundaries and morphological information using intelligent methods has become an important research direction in agricultural informatization and intelligent sensing. Compared to traditional manual interpretation methods, deep learning-based automated methods can improve processing efficiency in large-scale scenarios. However, in complex terrain backgrounds, weak boundary areas, and irregularly shaped regions, problems such as boundary discontinuities, unstable structural representations, and insufficient morphological regularity still easily arise.

[0003] Existing methods for extracting farmland information have generally evolved from traditional feature-based methods to deep learning methods. Early methods primarily relied on spectral and textural features of images and their combinations, such as threshold segmentation, unsupervised clustering, and supervised classification. These methods are applicable to simple scenes, but in situations with mixed land cover types, significant background interference, and obvious gradual transitions at plot edges, they often struggle to balance boundary localization accuracy with regional structural integrity. While traditional edge detection operators can extract local gradient change information, they are prone to producing false edges, broken edges, and jagged boundaries when faced with complex textures, field roads, ditches, and shadows in high-resolution scenes, making it difficult to meet the requirements of continuous and morphologically regular representation of farmland boundaries.

[0004] With the development of convolutional neural networks, researchers have proposed deep networks such as UNet, SegNet, ResUNet, RefineNet, and ERFNet for parcel extraction and boundary representation tasks. These methods have achieved some success in improving region representation accuracy and parcel integrity, but they still have shortcomings in handling boundary corners, elongated regions, weak boundary regions, and morphological continuity and structural consistency in complex scenes. Furthermore, some methods attempt to improve boundary response capabilities through multi-scale edge branching, attention mechanisms, or edge enhancement strategies. For example, DexiNed emphasizes multi-scale edge reconstruction, ResUNet-a utilizes residual structures to enhance information transmission, REAUNet improves boundary-related responses through edge detection modules and dual attention mechanisms, and BSNet uses region information to suppress pseudo-edges. Although the above methods have made improvements in boundary enhancement, region extraction, or pseudo-edge suppression, existing networks generally still have the following problems: First, they lack sufficient collaborative modeling of orientation-related features, boundary-related features, and region structural features, making it difficult to simultaneously ensure boundary representation accuracy and morphological integrity. Secondly, the cross-layer feature interaction and multi-scale fusion mechanisms are insufficient, which can easily lead to boundary breaks, local false responses, or structural discontinuities in complex scenarios. Thirdly, existing training methods tend to focus on local classification or local response enhancement, and do not adequately constrain the consistency between regional structural representation and boundary representation, thus making it difficult to effectively guarantee the uniformity of output results in terms of structural closure, morphological regularity, and overall stability.

[0005] Therefore, there is an urgent need to provide a new neural network architecture to improve the joint representation capability of farmland boundary and morphological information in complex scenarios, so as to obtain output results with more complete structure, more stable boundaries and more consistent representation. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a neural network architecture oriented towards boundary and morphological representation, including: a boundary representation branch, a shape semantic modeling branch, and a feature fusion branch, wherein: The boundary representation branch is used to encode the input feature tensor in multiple scale directions based on multiple directional geometric modules to generate a boundary representation tensor. The shape semantic modeling branch is used to perform hierarchical context modeling and structural representation enhancement on the input feature tensor based on multiple convolutional residual units and multiple morphological space-aware adaptive units, so as to generate a region structure representation tensor. The feature fusion branch is used to perform cross-layer interaction, weight modulation, and two-stream combination processing on the boundary representation tensor and the region structure representation tensor to generate the target output feature tensor.

[0007] In one possible implementation, the boundary representation branch includes: a preprocessing unit, an directional geometric representation unit, and a multi-scale boundary fusion module (MEFM); wherein: the preprocessing unit is used to perform an initial convolutional transformation on the input feature tensor to generate initial boundary features; the directional geometric representation unit is used to perform stepwise downsampling and directional response enhancement on the initial boundary features to generate boundary feature tensors of multiple scales; and the MEFM is used to perform scale alignment, channel aggregation, and feature reconstruction on the boundary feature tensors of multiple scales to generate boundary representation tensors.

[0008] In one possible implementation, the directional geometry representation unit includes multiple directional geometry sub-modules (DGMs). These DGMs decouple and model the directional information in the input features by introducing global convolution, horizontal convolution, and vertical convolution into the network, thereby enhancing the neural network's ability to represent boundary-related responses. The outputs of each branch are aggregated to form a corresponding level of directional enhancement feature tensor. in: Indicates input features, This is a global convolution used to extract the overall structural response. Horizontal convolution is used to extract horizontal response features. Vertical convolution is used to extract vertical response features.

[0009] In one possible implementation, the MEFM is used to perform scale alignment, channel aggregation, and feature reconstruction on boundary feature tensors at multiple scales to generate boundary representation tensors, including: After the boundary feature tensors of different scales are input into MEFM, the features at each level are first aligned by convolution, and then uniformly mapped to a common channel space to obtain aligned features. : in: For the first Tensor feature boundaries at various scales It is a linear rectification activation function; After obtaining the alignment features at four scales, respectively... , and Upsampling and alignment to a preset structural scale are performed, and the concatenated data is then combined with convolutional channels before feature reconstruction is performed to complete the fusion of multi-scale boundary representations. in: Indicates upsampling alignment. Indicates to conduct Double upsampling aligns features at each level to the preset structural scale. This is a splicing operation along the channel dimension. This is to compress and splice redundant channels; by As input, after Dropout regularization and After convolution to extract refined features, a multi-channel boundary representation output tensor is generated through Sigmoid activation. : in, express function, Boundary characterization results for each category; Further through Convolution compresses the number of channels to 1, and then normalizes it using a sigmoid function to obtain the final boundary representation output tensor: This is the final output of the boundary representation branch, which will be sent to the subsequent fusion module along with the region structure representation results of the shape semantic modeling branch.

[0010] In one possible implementation, the shape semantic modeling branch includes an encoding path, a context aggregation path, and a decoding path. The encoding path consists of multiple basic convolutional residual blocks (BCBlocks) used to extract hierarchical features and maintain gradient propagation stability. The context aggregation path includes a dilated pyramid pooling module (ASPP), which is used to extract and aggregate feature responses at different dilation rates in parallel to enhance the shape semantic modeling branch's ability to represent multi-scale structural relationships and global contextual information. The decoding path is used to perform feature reconstruction step by step and output a region structure representation tensor.

[0011] In one possible implementation, the shape semantic modeling branch further includes a morphological space-aware adaptive block FSAAB, which integrates morphological attention and spatial attention mechanisms and achieves the optimal combination of dual-path information through an adaptive weight allocation strategy. The morphological attention is used to model the response intensity of features across different channel dimensions. First, global average pooling and max pooling are performed on the input features to obtain two channel description vectors. These vectors are then processed by two fully connected networks with shared weights and summed. Finally, the vectors are normalized using a sigmoid activation function to form a channel attention map. : The spatial attention is used to model the spatial location distribution of key structural regions. Average pooling and max pooling operations are performed on the input features along the channel dimension, and the concatenated features are then processed through a convolutional layer and a sigmoid activation layer to generate a spatial attention map. : A lightweight weight generator is introduced to dynamically generate two normalized fusion coefficients based on the input global context information, which are used to adjust the strength of channel attention and spatial attention respectively: Finally, the fused output is weighted and corrected using channel attention and spatial attention on the original features: in: It is a channel attention network consisting of two fully connected layers. for, for, This indicates that the channel features are averaged and pooled. This indicates that max pooling is performed on the channel features. This indicates element-wise multiplication. For the adaptive fusion coefficients of the channel attention path, For the adaptive fusion coefficients of the spatial attention path, This is the fusion output feature tensor of FSAAB.

[0012] In one possible implementation, the feature fusion branch includes a cross-layer fusion unit and a two-stream combination unit; wherein, the cross-layer fusion unit is used to receive decoded features from the shape semantic modeling branch, corresponding scale features from the boundary representation branch, and the fusion result of the previous layer, and to sum or concatenate the decoded features, corresponding scale features, and the fusion result of the previous layer to form a feature tensor to be fused; the two-stream combination unit is used to perform orientation structure reshaping and channel recalibration on the feature tensor to be fused to generate the current layer fused feature tensor.

[0013] In one possible implementation, the dual-stream combination unit includes multiple multi-layer dual-stream combination modules (DSCMs). The DSCMs couple the synthesized features at different scales layer by layer, utilizing a squeeze excitation module (SE) and a channel attention module (SM) to highlight channel features that contribute to the task and suppress redundant responses. Vertical and horizontal convolutional channels are introduced again to reorganize multi-scale features in the cross-direction, thus more fully modeling the directional dependencies of irregular structures. in: For input features, For extrusion excitation module, For channel attention module, This is a multi-scale feature in the vertical direction obtained by concatenating vertically dilated convolutions at four dilation rates. This is a multi-scale feature in the horizontal direction obtained by concatenating horizontally dilated convolutions at four dilation rates. This is a global oriented multi-scale feature obtained by concatenating global dilated convolutions at four dilation rates. Indicates splicing, , and Represents vertical, horizontal, and global convolution. Indicates the expansion rate; At each layer, the decoded features from the shape semantic modeling branch, the corresponding scale features from the boundary representation branch, and the fusion features generated in the previous layer are additively fused and then input into the DSCM module to enhance the representation of orientation and scale structure. Subsequently, feature reconstruction is performed through transposed convolution, and the number of feature channels is compressed to a smaller dimension to complete the feature reconstruction. Convolution is then used to further smooth and fuse features, suppressing redundant responses introduced during feature reconstruction. Finally, convolution is mapped to the number of categories to obtain the region structure output tensor. in: For the first The fusion output feature tensor of the layers For the first Feature transformation operations of the layered dual-stream combination module Decoding path for shape semantic modeling branch Decoding features of layer output, For the boundary characterization branch Boundary features corresponding to the scale of the layer. For hierarchical restructuring operations, This is a transposed convolution; After obtaining the region structure output tensor, it is jointly encoded with the boundary representation output tensor. Then, the joint representation is reorganized using symmetric feature reconstruction and channel compression operations to generate the final target output feature tensor.

[0014] In one possible implementation, a loss including region representation is constructed. Category equilibrium boundary loss and region-boundary consistency constraint loss Multi-task joint training objective function : in, , and These are the weight coefficients of the corresponding loss function; in: This represents the total number of basic representation units involved in the loss calculation. Indicates the number of categories. For the first The real class label encoding vector corresponding to the basic representation unit is in the _th _ Components on class, For the first The fundamental representation unit in the th... The predicted output value of the class; in, and These represent the set of real boundary representation units and the set of non-boundary representation units, respectively. The boundary representation value output by the boundary representation branch. and This is a dynamically calculated class balance coefficient used to enhance the constraint effect of sparse boundary samples on parameter updates during backpropagation.

[0015] In one possible implementation, By using a differentiable local structure difference operator, difference extraction is performed on the output tensor of the region structure representation to obtain the structural difference representation: In the formula, This represents the convolution operation. Used to truncate outliers after differentiation and normalize them. interval, The output tensor representing the region structure from the shape semantic modeling branch has a differentiable structure difference kernel as follows: After obtaining the structural difference representation, the final consistency constraint term is constructed by calculating the norm distance between it and the output of the boundary representation branch: in: To perform difference extraction on the output tensor representing the region structure using a differentiable local structure difference operator, the th... Structural difference representation values ​​of the locations of the basic representation units For boundary characterization branches in the th The boundary representation output value of each basic representation unit location. Represents norm operations.

[0016] Secondly, embodiments of this application provide a system for boundary and morphological representation, which is equipped with a neural network architecture as described in any possible implementation of the first aspect, for extracting field boundaries.

[0017] In this embodiment, the boundary representation branch encodes the input feature tensor using multi-scale directional responses based on multiple directional geometric modules, enhancing the network's ability to perceive weak boundaries, complex contours, and multi-directional boundary information, thereby improving the continuity and stability of boundary representation. The shape semantic modeling branch performs hierarchical context modeling and structural representation enhancement based on multiple convolutional residual units and multiple morphological space-aware adaptive units, improving the ability to express the overall morphology, structural relationships, and contextual information of the region. The feature fusion branch further performs cross-layer interaction, weight modulation, and dual-stream combination processing on the boundary representation tensor and the region structural representation tensor, improving the fusion effect between features at different levels and features of different types. Therefore, it effectively alleviates problems such as boundary breaks, local erroneous responses, structural discontinuities, and insufficient morphological regularity in existing technologies, resulting in output results with better structural integrity, boundary stability, and representation consistency. Attached Figure Description

[0018] Figure 1 A schematic diagram of a neural network architecture for boundary and morphological representation provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of the DMG provided in the embodiments of this application; Figure 3 A schematic diagram of MEFM provided for an embodiment of this application; Figure 4 This is a schematic diagram of the structure of FSAAB provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of the DSCM provided in the embodiments of this application; Figure 6 The false edges and breaks shown in the results provided by the embodiments of this application; Figure 7 This is a schematic diagram illustrating the model test comparison results provided in the embodiments of this application; Figure 8 A schematic diagram comparing model accuracy and model storage overhead provided for embodiments of this application; Figure 9 A schematic diagram comparing model accuracy and model parameter scale provided in the embodiments of this application; Figure 10 A schematic diagram comparing model accuracy and floating-point computational complexity provided in this application embodiment; Figure 11 A schematic diagram comparing model accuracy and model inference time provided for embodiments of this application; Figure 12 This is a schematic diagram of the ablation experiment visualization results provided in the embodiments of this application; Figure 13 This is a schematic diagram comparing the results of FPESNet field edge extraction at different spatial resolutions provided in the embodiments of this application; Figure 14 This is a schematic diagram comparing field edge feature maps under different parameters provided in the embodiments of this application; Figure 15 This is a schematic diagram of the framework of a system for boundary and morphological representation provided in an embodiment of this application. Detailed Implementation

[0019] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0020] See Figure 1 This embodiment provides FPESNet, a neural network architecture oriented towards boundary and morphological representation. For example... Figure 1 As shown, FPESNet mainly consists of a boundary representation branch, a shape semantic modeling branch, and a feature fusion branch. The boundary representation branch utilizes DGM (Directional Geometry Model) to enhance directionally relevant features and generate a boundary representation tensor. The shape semantic modeling branch employs BCBlock (Basic Conv Block) and FSAAB (Form-Spatial Aware Adaptive Block) from REAUNet for hierarchical context modeling and structural representation enhancement to obtain regional structural representations. The feature fusion branch further performs cross-layer interaction, weight modulation, and two-stream combination processing on the regional structural representation and boundary representation to generate a target output feature tensor that balances structural integrity and boundary consistency.

[0021] In this embodiment, the boundary representation branch includes: a preprocessing unit, an directional geometry representation unit, and a MEFM (Multi-scale Edge Fusion Module). Specifically: the preprocessing unit performs an initial convolutional transformation on the input feature tensor to generate initial boundary features. The directional geometry representation unit performs progressive downsampling and directional response enhancement on the initial boundary features to generate boundary feature tensors of multiple scales. The MEFM performs scale alignment, channel aggregation, and feature reconstruction on the boundary feature tensors of multiple scales to generate a boundary representation tensor.

[0022] To address the characteristics of significant differences in field scale, complex directional variations, and irregular local contours, this framework introduces a direction-sensitive Directional Generalized Mechanism (DGM), the structure of which is as follows: Figure 2 As shown, multi-scale directional response encoding is performed on the input feature tensor to enhance the network's ability to represent boundary information in different directions and weak boundary information: in: Indicates input features, This is a global convolution used to extract the overall structural response. Horizontal convolution is used to extract horizontal response features. Vertical convolution is used to extract vertical response features. The above formula shows that DGM decouples and models the directional information in the input features by introducing global convolution, horizontal convolution, and vertical convolution into the network, thereby enhancing the neural network's ability to represent boundary-related responses. The outputs of each branch are aggregated to form the corresponding layer's directional enhancement feature tensor.

[0023] Meanwhile, to enhance the structural expressiveness at different scales, preliminary features are first extracted through convolution with a stride of 2, then refined through a DGM with a stride of 1, and preprocessed features are obtained through upsampling alignment and feature reconstruction. Subsequently, DGM with a stride of 2 is used to progressively downsample, generating a four-level multi-scale feature sequence, as shown in the formula: in, This indicates an upsampling alignment or feature reconstruction operation. This indicates that the output step size of the DGM is 2.

[0024] Multi-scale boundary representation is particularly important for improving boundary continuity and structural consistency. Therefore, design and implement MEFM (Mean Scale Meaning). See [link to MEFM]. Figure 3 The boundary feature tensors at different scales are fed into the MEFM module. First, convolution is used to align the channels of the features at each level, mapping them uniformly to a common channel space to obtain aligned features. : in: For the first Tensor feature boundaries at various scales It is a linear rectification activation function.

[0025] After obtaining the alignment features at four scales, respectively... , and Upsampling and alignment to a preset structural scale are performed, and the concatenated data is then combined with convolutional channels before feature reconstruction is performed to complete the fusion of multi-scale boundary representations. in: Indicates upsampling alignment. Indicates to conduct Double upsampling aligns features at each level to the preset structural scale. This is a splicing operation along the channel dimension. This is for the redundant channels after compression and splicing.

[0026] by As input, after Dropout regularization and After convolution to extract refined features, a multi-channel boundary representation output tensor is generated through Sigmoid activation. : in, express function, Boundary representation results for each category.

[0027] Further through Convolution compresses the number of channels to 1, and then normalizes it using a sigmoid function to obtain the final boundary representation output tensor: This is the final output of the boundary representation branch, which, together with the region structure representation results of the shape semantic modeling branch, will be fed into the subsequent fusion module to enhance the boundary consistency and structural stability of the network output.

[0028] Due to the significant differences in morphology, scale, and spatial organization of field plots, and the rich local details and contextual information contained in high-resolution data, traditional shallow convolutional networks struggle to effectively represent multi-scale structures. To construct an efficient deep network and alleviate the gradient vanishing problem, this embodiment utilizes BCBlock from REAUNet and the proposed FSAAB to perform feature aggregation at the morphological and channel feature levels, obtaining multi-level regional structure representations. The encoder and decoder employ the Aperture Pyramid Pooling (ASPP) module for multi-scale feature fusion, thereby acquiring global contextual information.

[0029] In this embodiment, the shape semantic modeling branch includes an encoding path, a context aggregation path, and a decoding path. The encoding path consists of multiple basic convolutional residual blocks (BCBlocks) used to extract hierarchical features and maintain gradient propagation stability. The context aggregation path includes a dilated pyramid pooling module (ASPP), which is used to extract and aggregate feature responses at different dilation rates in parallel, enhancing the shape semantic modeling branch's ability to represent multi-scale structural relationships and global contextual information. The decoding path is used to perform feature reconstruction step-by-step and output a region structure representation tensor.

[0030] join Figure 4 In this embodiment, FSAAB integrates morphological attention and spatial attention mechanisms, and achieves the optimal combination of dual-path information through an adaptive weight allocation strategy, thereby improving the model's ability to perceive key structural regions. Morphological attention is used to model the response intensity of features in different channel dimensions. First, the input features are subjected to global average pooling and max pooling to obtain two channel description vectors. Then, these vectors are processed by two fully connected networks with shared weights and added together. Finally, the channel attention map is formed by normalization using the Sigmoid activation function. : The spatial attention is used to model the spatial location distribution of key structural regions. Average pooling and max pooling operations are performed on the input features along the channel dimension, and the concatenated features are then processed through a convolutional layer and a sigmoid activation layer to generate a spatial attention map. : To balance the contributions of the two attention paths, a lightweight weight generator is introduced to dynamically generate two normalized fusion coefficients based on the global context information of the input. , These are used to adjust the intensity of channel attention and spatial attention, respectively. Finally, the fused output uses channel attention and spatial attention to weighted modify the original features, thereby enhancing the representation of the region structure. in: It is a channel attention network consisting of two fully connected layers. for, for, This indicates that the channel features are averaged and pooled. This indicates that max pooling is performed on the channel features. This indicates element-wise multiplication. For the adaptive fusion coefficients of the channel attention path, For the adaptive fusion coefficients of the spatial attention path, This is the fusion output feature tensor of FSAAB.

[0031] In this embodiment, the feature fusion branch includes a cross-layer fusion unit and a two-stream combination unit. The cross-layer fusion unit receives decoded features from the shape semantic modeling branch, corresponding scale features from the boundary representation branch, and the fusion result from the previous layer. It then sums or concatenates the decoded features, corresponding scale features, and the fusion result from the previous layer to form a feature tensor to be fused. The two-stream combination unit performs orientation restructuring and channel recalibration on the feature tensor to be fused to generate the current layer fused feature tensor.

[0032] The comprehensive features derived from each layer of the boundary representation branch and shape semantic modeling branch can effectively represent directional, boundary, and regional structural information at different scales. However, the spatial arrangement and scale distribution of fields vary significantly in different scenarios. Therefore, this embodiment introduces a multi-layer dual-stream composition module (DSCM) to aggregate the comprehensive features at different scales layer by layer, such as... Figure 5 As shown.

[0033] In this embodiment, the dual-stream combination unit includes multiple multi-layer dual-stream combination modules (DSCMs). These DSCMs couple integrated features at different scales layer by layer, utilizing a squeeze excitation module (SE) and a channel attention module (SM) to highlight channel features that contribute to the task and suppress redundant responses. Vertical and horizontal convolutional channels are introduced again to reorganize multi-scale features in the intersecting directions, providing a more comprehensive model of the directional dependencies of irregular structures. The use of multi-layer DSCMs helps improve the network's joint representation ability and overall robustness to complex structures. in: For input features, For extrusion excitation module, For channel attention module, This is a multi-scale feature in the vertical direction obtained by concatenating vertically dilated convolutions at four dilation rates. This is a multi-scale feature in the horizontal direction obtained by concatenating horizontally dilated convolutions at four dilation rates. This is a global oriented multi-scale feature obtained by concatenating global dilated convolutions at four dilation rates. Indicates splicing, , and Represents vertical, horizontal, and global convolution. Indicates the expansion rate; To fully integrate the complementary information of the shape semantic modeling branch and the boundary representation branch at different scales, and to enhance the consistent representation of the network output, a stepwise residual fusion path was designed. At each layer, the decoded features of the shape semantic modeling branch, the corresponding scale features of the boundary representation branch, and the fusion features generated in the previous layer are additively fused and then input into the DSCM module to enhance the representation of orientation and scale structure. Subsequently, a feature reconstruction operation is performed through transposed convolution, compressing the number of feature channels to a smaller dimension to complete the feature reconstruction. Convolution is then used to further smooth and fuse the features, suppressing redundant responses introduced during feature reconstruction. Finally, convolution is used to map to the number of categories to obtain the region structure output tensor. in: For the first The fusion output feature tensor of the layers For the first Feature transformation operations of the layered dual-stream combination module Decoding path for shape semantic modeling branch Decoding features of layer output, For the boundary characterization branch Boundary features corresponding to the scale of the layer. For hierarchical restructuring operations, This is a transposed convolution; After obtaining the region structure output tensor, it is jointly encoded with the boundary representation output tensor. Next, to expand the receptive field and reduce information loss during downsampling, two dilated convolutional layers with a dilation rate of 2 are used for further feature extraction. Subsequently, symmetric feature reconstruction and channel compression operations are used to reorganize the joint representation, generating the final target output feature tensor, achieving the coordinated output of boundary representation and region structure representation.

[0034] To simultaneously improve the consistency of region structure representation and the stability of boundary representation, this embodiment designs a method that includes region representation loss. Category equilibrium boundary loss and region-boundary consistency constraint loss Multi-task joint training objective function : in, , and These are the weight coefficients of the corresponding loss functions, serving as hyperparameters to balance the influence of the training objective function.

[0035] For the region structure representation task, the basic representation unit-level cross-entropy loss function is used to supervise the class prediction and ground truth labels of the network output. The formulas for calculating the region representation loss and boundary loss are as follows: in: This represents the total number of basic representation units involved in the loss calculation. Indicates the number of categories. For the first The real class label encoding vector corresponding to the basic representation unit is in the _th _ Components on the class, For the first The fundamental representation unit in the th... The predicted output value of the class.

[0036] in, and These represent the set of real boundary representation units and the set of non-boundary representation units, respectively. The boundary representation value output by the boundary representation branch. and This is a dynamically calculated class balance coefficient used to enhance the constraint effect of sparse boundary samples on parameter updates during backpropagation.

[0037] In conventional multi-task joint frameworks, region representation loss and boundary loss are typically calculated independently. This mechanism can easily lead to the region structure representation output and boundary representation output becoming separated and offset in the feature space during network training. To explicitly enhance the consistency of the two-branch outputs mathematically, this embodiment proposes a region-boundary consistency constraint loss. .

[0038] In this embodiment By using a differentiable local structure difference operator, difference extraction is performed on the output tensor of the region structure representation to obtain the structural difference representation: In the formula, This represents the convolution operation. Used to truncate outliers after differentiation and normalize them. interval, The output tensor representing the region structure from the shape semantic modeling branch has a differentiable structure difference kernel as follows: After obtaining the structural difference representation, the final consistency constraint term is constructed by calculating the norm distance between it and the output of the boundary representation branch: in: To perform difference extraction on the output tensor representing the region structure using a differentiable local structure difference operator, the th... Structural difference representation values ​​of the locations of the basic representation units For boundary characterization branches in the th The boundary representation output value of each basic representation unit location. Represents norm operations.

[0039] During gradient backpropagation, the network is required to ensure that the locations of abrupt changes in the regional structural representation and the high-response locations of the boundary representation branches remain consistent in the representation space. This explicit consistency constraint helps alleviate problems such as structural discontinuities, local erroneous responses, and boundary instability, thereby improving the network's ability to jointly represent boundary and morphological information in complex scenarios.

[0040] The network architecture provided in this embodiment is implemented using the PyTorch framework and trained and tested using a single NVIDIA GeForce RTX 3060 GPU. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, and performance evaluation is based on the validation set results. To enhance the diversity of training data and alleviate overfitting, data augmentation strategies such as contrast adjustment, brightness adjustment, and multi-angle random flipping are employed in the preprocessing stage. The experimental environment configuration is as follows: PyTorch version 2.1.0, CUDA version 12.4, and core dependencies including Python 3.9.13, GDAL 3.6.0, NumPy 1.26.4, OpenCV-Python 4.10.0.84, and Scikit-learn 1.0.2. The model is optimized using a multi-task joint training objective function, and the Adam optimizer is used for parameter updates. The input feature tensor size is 512×512. The training hyperparameters are set as follows: batch size of 6, number of iterations of 120, and learning rate of 0.0005. After each training round, a model parameter file is generated for evaluation, and the set of parameters with the best objective function value on the validation set is selected as the best model to test its performance.

[0041] To verify the performance of the aforementioned neural network architecture, this application also conducted a series of experimental verifications.

[0042] This embodiment uses remote sensing imagery of land parcels for validation, but other types of data besides images can also be used. To evaluate the applicability and robustness of the proposed network under different regional conditions and remote sensing imaging conditions, this embodiment selects four counties in Shandong Province, China, as the study area. In terms of regional representativeness, the study area covers typical agricultural areas of Shandong. As a major agricultural province in China, Shandong is highly representative in agricultural remote sensing monitoring and farmland protection research. The selected area is located in the Huang-Huai-Hai Plain and its transition zone, with diverse landforms, including typical combinations of plains, hills, and waterways. It includes Dongping County in Tai'an City, Huantai County and Gaoqing County in Zibo City, and Dong'e County in Liaocheng City.

[0043] Regarding edge extraction difficulty, these four counties provide complex test scenarios with rich gradients. In Dongping County, Tai'an City, some areas are characterized by irregular and fragmented field distribution due to hilly terrain and complex water systems, resulting in complex, curved boundaries and dramatic scale variations. Such extreme scale and morphological changes easily cause existing models to break down or fail locally during edge extraction. While Huantai County, Gaoqing County, and Dong'e County are predominantly plains with relatively continuous boundaries, they also suffer from background features and weak field edges, among other interfering factors. Therefore, the areas selected in this embodiment are not only highly representative of national agriculture but also present extremely high extraction challenges in terms of terrain and morphological features, enabling rigorous and comprehensive verification of the effectiveness of the neural network architecture in different scenarios.

[0044] The data used in the verification experiment of this embodiment mainly includes Gaofen-2 1m resolution imagery and Beijing-2 0.5m and 0.8m resolution imagery. The multi-source, multi-resolution high-resolution imagery not only ensures the integrity of spatial details and the richness of spectral information, but also provides multi-level verification conditions for the field edge extraction algorithm. Comparative analysis of imagery at different resolutions can effectively evaluate the adaptability and stability of the edge detection algorithm at different spatial scales; while the combination of samples across regions and across plot morphologies further enhances the robustness and universality of the research conclusions.

[0045] This study selected GF2 imagery provided by the China Center for Resources Satellite Data and Applications, which was then fully stitched together to cover the study area. Based on an open-source remote sensing image processing platform, a standardized preprocessing workflow was implemented on the raw data: first, radiometric calibration was performed to eliminate sensor response differences; then, atmospheric correction was used to eliminate aerosol effects; orthorectification was performed using a digital elevation model to eliminate topographic distortion; finally, pixel-level fusion technology was used to fuse panchromatic and multispectral data to generate a four-band composite image with a spatial resolution of 1 meter. The preprocessed data provided a high-quality, standardized remote sensing dataset for subsequent analysis. The BJ-2 imagery used the standard preprocessing product released by the data provider, which had already undergone radiometric, atmospheric, and geometric corrections during the production phase, and provided a fused image at the corresponding resolution.

[0046] The dataset primarily includes land cover types such as buildings, roads, water bodies, woodlands, bare land, wheat fields, and other crop land. Labels were created on the remote sensing images through visual interpretation, using 1 for crop land pixels and 0 for non-crop land pixels. Edge labels used 1 for crop land edges and 0 for non-crop land edges. After labeling, the remote sensing images and their corresponding labeled images were segmented into 512×512 pixel image blocks, resulting in 1800 pairs of "remote sensing image-labeled image." These pairs formed the training dataset for training and testing the model.

[0047] Based on the multi-source high-resolution remote sensing image data, including images contained in the aforementioned dataset, the design and validation were conducted. The remote sensing images of farmland exhibit two significant characteristics: (1) morphological diversity, with significant differences in the size, shape, and orientation of the fields. (2) complex topological structure, with various spatial relationships such as mosaic, fragmentation, and nesting among the fields. These two characteristics often lead to problems such as incomplete outlines, blurred and discontinuous edges, and irregular shapes in the field extraction results, causing the model's prediction results to deviate from the actual field boundaries, thereby affecting the accuracy and efficiency of precision agriculture management.

[0048] To comprehensively evaluate the effectiveness of FPESNet, this paper selects two representative comparison models: one is the general edge detection network DexiNed, which only outputs edges; the other is a multi-task or boundary-enhanced semantic segmentation network that integrates boundary information, including BESNet, REAUNet, ResUNet-a, and BSNet. All the comparison models are implemented based on publicly available code and retrained on the dataset of this embodiment. The parts without publicly available code are implemented according to the original structure. All models use a uniform input size, data augmentation strategy, optimizer and learning rate plan, batch size, and number of training epochs on the image dataset of this embodiment, and are trained and evaluated on the same hardware platform to ensure fairness.

[0049] This embodiment uses five metrics—overall accuracy (OA), precision (P), recall (R), F1 score (F1), and intersection-over-union (IoU)—to evaluate the edge extraction test results. Edge predictions from all models are output as grayscale probability maps, with pixel values ​​ranging from [value range missing]. This can be viewed as a marginal confidence map. For ease of standardized evaluation, the predicted grayscale image is first normalized to... and adopt a uniform threshold Convert it into a binary edge mask and set the confidence level. Pixels that are considered edge pixels are assigned a value of 1, otherwise a value of 0, resulting in a predicted binary edge map. Correspondingly, the true labeled binary edge map is denoted as... .

[0050] Meanwhile, this embodiment introduces a tolerance distance during edge matching. (Unit: pixels), the distance is at An edge within the specified range is considered a successful match. Define the distance set operator: in: Represents pixels to the set The Euclidean distance to the nearest edge pixel. Equivalently, This can be understood as using the radius Structural elements at the edges The tolerance band is obtained by expansion.

[0051] within tolerance distance Down, , , and The definitions are as follows: in This represents the number of edge pixels in the set.

[0052] Accuracy is used to evaluate the classification performance of the model after thresholding the edge extraction results, i.e., the proportion of correctly predicted samples out of the total samples. Precision is used to evaluate the model's accuracy in predicting the target class, i.e., the proportion of samples predicted as belonging to the target class that actually belong to that class. Recall is used to evaluate the model's ability to identify the target class, i.e., the proportion of samples that actually belong to the target class that are correctly predicted. F1-score is used to evaluate the balance between precision and recall, measuring the model's overall performance in terms of precision and recall. IoU is used to evaluate the degree of overlap between the edge pixel set predicted by the model and the ground truth edge pixel set, directly reflecting the consistency of the field boundary position and shape.

[0053] The calculation method for OA, P, R, and F1 scores in terms of IoU is as follows: Among them, the true example ( ) represents the number of pixels correctly predicted as positive examples, and the number of false positive examples ( ) represents the number of pixels that incorrectly predicted positive examples, and the number of pixels that correctly predicted negative examples. ) represents the number of pixels correctly predicted as negative examples, and the number of false negative examples ( () indicates the number of pixels that are incorrectly predicted as counterexamples. This represents the number of true edge pixels used for Precision, i.e., the number of pixels predicted as edges that fall within the true edge. The number of pixels in the pixel neighborhood; This represents the number of true instance pixels used for recall, i.e., the number of pixels that, out of those actually labeled as edges, fall within the predicted edge. The number of pixels in the pixel's neighborhood. Correspondingly, This represents the number of pixels predicted as edges but not matching the true edges within the tolerance distance; This represents the number of pixels in the true edge that are not covered by the predicted edge within the tolerance distance.

[0054] Traditional deep learning models like BESNet and ResUNet-a have high response intensity to the internal texture and shadow details of fields, but they lack semantic constraints and are prone to misclassifying internal noise as field boundaries, resulting in low precision in problematic images such as... Figure 6 As shown in (a), FPESNet further enhances the semantic consistency of objects and the feature selection of boundary regions through FSAAB, making the edge response consistent with the overall field, suppressing scattered pseudo-edges and improving the smoothness and regularity of the boundaries, such as... Figure 6 As shown in (b), DexiNed, as a general network that only outputs edges, relies more on local gradient information. In scenes with weak contrast boundaries and small fields, it is prone to edge breakage and missed detection, resulting in low Recall and limited IoU accuracy. Problem images are shown below. Figure 6 As shown in (c), FPESNet introduces DGM and MEFM into its edge search branch. Through multi-scale fusion, it effectively expands the receptive field and enhances orientation awareness, addressing the problem of weak boundary fragmentation. DGM enhances the orientation sensitivity to boundaries with different orientations, enabling more stable capture of boundary cues such as weak contrast and narrow field ridges. MEFM expands the effective receptive field through multi-scale feature fusion and highlights stable boundary responses during the fusion process, thus simultaneously improving boundary integrity and discriminability from a mechanistic perspective. Figure 6 As shown in (d), while REAUNet and BSNet methods alleviate the false edge problem to some extent, they are still insufficient for multi-directional, multi-scale geometric modeling and hierarchical collaborative optimization of complex boundaries. Blurring may still occur in small plots and complex corners, as shown in the problematic image. Figure 6As shown in (e), FPESNet uses DSCM for hierarchical bidirectional semantic-edge fusion, enabling high-level semantics to constrain boundary localization and low-level details to compensate for contour refinement, reducing boundary blurring and breaks, and improving IoU accuracy. Figure 6 As shown in (f).

[0055] Table 1 Comparison of evaluation indicators for different models Table 1 presents the evaluation metrics for the experimental results of the dataset in the experimental area involved in this study. Five typical results were selected from the experimental results: plots containing buildings (A), regularly shaped plots (B), strip-shaped plots (C), and plots containing scattered plots and hilly areas (D, E). These results are used to visually demonstrate the extraction performance of the FPESNet method and the comparison model. Figure 7 As shown in the figure. The first column is the remote sensing image, and the second to eighth columns are the extraction results and ground truth labels of the FPESNet, REAUNet, DexiNed, BESNet, ResUNet-a, and BSNet models, respectively.

[0056] In edge extraction tasks, BESNet, by treating edge extraction as a unidirectional auxiliary task to improve semantic consistency, generates edges with poor continuity and irregular shapes, resulting in false edges. BESNet and ResUNet-a models, lacking semantic segmentation information to constrain edge extraction, exhibit noticeable false edges in the target field regions, as shown in columns B-6 and B-7, and C-6 and C-7. REAUNet and BSNet models incorporate semantic segmentation information to assist edge extraction, improving field integrity and shape and reducing false edges, but a small number still exist, as shown in columns B-4 and B-8, and C-4 and C-8. The DexiNed model uses dense skip connections to learn from scratch and preserve sharp edges for edge extraction. While reducing false edges, this leads to missed edges in small field regions, as shown in columns C-5 and D-5. In hilly areas, when faced with small and fragmented fields, BESNet and ResUNet-a exhibit false edges within the fields. While DexiNed does not exhibit false edges within the fields, it shows incomplete edges in some areas. The FPESNet model proposed in this embodiment performs relatively well, with no obvious false edges within the fields. Overall, the FPESNet model demonstrates the best performance in all test cases: firstly, by using semantic segmentation information to constrain edge extraction, it produces smooth edges for large fields, reflecting the structural integrity of the fields; secondly, multi-scale information fusion provides more accurate contour extraction for small fields.

[0057] Figures 8-11 This is a visualization comparing the efficiency of the comparison models ResUNet-a, REAUNet, and BSNet. Compared to traditional segmentation models, ResUNet-a, REAUNet, and BSNet are representative models with high accuracy, but these models require huge computational overhead, which limits their application in resource-constrained scenarios. ResUNet-a, in particular, stacks standard residual convolutions across its extremely deep backbone network and relies on parallel multi-scale dilated convolutions for multi-task decoding, leading to a surge in matrix operations. REAUNet has a parameter size as high as 142.40MB, FLOPs of 254.42G, and an inference time as long as 0.25s. This is because REAUNet embeds a complex dual attention mechanism and explicit edge detection module into its traditional encoding and decoding structure; this cascading stacking significantly increases the network's parameter burden. BSNet also employs a computationally complex fusion strategy to supplement semantic segmentation information, resulting in FLOPs as high as 357.27G. In contrast, FPESNet has only 10.55MB of parameters and 79.73G of FLOPs. FPESNet extensively employs depthwise separable convolutions in its semantic branch, significantly reducing the number of parameters and computational cost when extracting local spatial features from each channel. Meanwhile, the morpho-spatial-aware adaptive block (FSAAB) dynamically allocates attention coefficients through a lightweight weight generator, avoiding the introduction of a large fully connected layer network and achieving better computational efficiency while maintaining feature representation capabilities.

[0058] Compared to lightweight models, DexiNed and BESNet have advantages in parameter size and inference speed, but at the expense of some completeness and accuracy in field edge extraction. DexiNed has 35.08MB of parameters and an inference time of 0.16s, but its IoU is only 0.7884. In principle, DexiNed uses dense skip connections to learn the edge map directly from the bottom layer, relying too much on local gradient information and lacking global object-level semantic constraints, which makes it prone to edge breakage and missed detection in complex farmland scenarios. Although BESNet introduces semantic information, it only treats boundary extraction as a unidirectional auxiliary task to improve semantic consistency, lacking deep bidirectional feature fusion, resulting in poor output edge continuity and an IoU of only 0.7928. Although FPESNet's inference time is slightly higher than the other two, it achieves the best extraction accuracy. Mechanistically, FPESNet designs a multi-layer dual-stream combinatorial module (DSCM), which uses squeezing excitation and channel attention to perform layer-by-layer, bidirectional fusion of comprehensive features at different scales during the decoding stage. This mechanism compensates for the lack of global context awareness in lightweight backbone networks with minimal computational cost, effectively suppresses pseudo edges and fills in breaks, achieving a leap in accuracy and efficiency.

[0059] Overall, FPESNet does not pursue extreme minimization in all dimensions. Instead, it achieves a comprehensive capability that better meets the needs of agricultural field extraction tasks by striking a balance between parameter size, computational complexity, and accuracy. In real-world applications where accuracy, speed, and resource consumption must be considered, FPESNet has greater practical potential.

[0060] To verify the contribution of key modules in FPESNet to the performance of field edge extraction, this embodiment designed a systematic ablation experiment on the constructed dataset to examine the roles of the multi-scale edge fusion module (MEFM), the two-stream combination module (DSCM), and the morphology-space-aware adaptive block (FSAAB).

[0061] The ablation experiments used a U-Net architecture as the baseline model: the encoder used ResNet-18, the decoder used layer-wise transposed convolutional upsampling, and the field segmentation results were output with a single semantic segmentation branch and a Sigmoid classifier, without any edge branches or attention modules. Based on this, MEFM, DSCM, and FSAAB were introduced in sequence to form various network configurations.

[0062] Figure 12 Visual comparison results of the ablation experiments are presented. As can be seen from the figures, compared to the baseline, adding MEFM to the network significantly improves the accuracy of field edges and reduces the coarseness of the edge results, as shown in the comparison between rows A and B in the third and fourth columns. Further introducing DSCM, using the semantic segmentation results to constrain the edge extraction of field contours results in more regular shapes, and the boundaries of irregular fields are also more finely characterized, as shown in the comparison between rows A, B, and D in the fourth and fifth columns. When FSAAB is simultaneously introduced, i.e., the complete FPESNet model, the introduction of a dual adaptive attention mechanism of channels and space improves the accuracy of semantic segmentation results. Not only are the field edges more complete, but the noise within small fields is also significantly reduced, and the field areas are cleaner, as shown in rows C and D in the sixth column. To analyze the effectiveness of the introduced branches on the model, the performance metrics comparison results of the ablation experiments are shown in Table 2.

[0063] Table 2 Comparison of Evaluation Indicators for Ablation Experiment Results As shown in Table 2, after introducing the MEFM module into the baseline model, the F1 and IoU indices improved by 1.54% and 2.54%, respectively. This change indicates that the MEFM module enhances the edge information representation effect by cross-scale cascading and fusing directional edge responses of different depth levels, effectively improving the model's ability to discriminate target boundary regions. Further introducing the DSCM module, using multi-dilation rate convolution and attention mechanisms, injects high-level semantic consistency priors into edge features, resulting in a slight improvement in IoU and F1. Finally, introducing the FSAAB module, through a lightweight weight generator and adaptive channel allocation and spatial attention mechanisms, achieves precise focusing on key regions during the feature encoding stage, enabling the model's OA, Precision, Recall, F1, and IoU to reach 95.56%, 88.91%, 98.21%, 93.33%, and 87.49%, respectively, further improving edge extraction accuracy. Compared to the baseline, this combination achieves improvements in evaluation metrics, particularly in edge consistency and region overlap. This result further verifies the cumulative effect brought about by the collaborative interaction of the three modules, which helps to enhance the robustness and generalization ability of the model in complex scenarios.

[0064] To analyze the adaptability and robustness of FPESNet under different spatial resolutions, this embodiment conducted a multi-resolution comparative experiment based on the original data. In addition to 0.5m, 0.8m, and 1m resolution imagery, remote sensing imagery at 0.3m and 2m resolutions was further added. FPESNet was then trained and tested at each resolution to systematically evaluate its performance trends under multi-scale conditions.

[0065] Depend on Figure 13 As can be seen, at a resolution of 0.3m, FPESNet can still accurately delineate the boundaries of major fields. However, due to the extremely rich detail and drastic variations in field color and texture, intra-class differences within fields increase significantly. High-frequency textures such as local operation traces, small ditches, and shadows are treated as potential boundaries by the model in a few regions, thus introducing a certain number of false edges and local breaks. The root cause of this failure is that the model's lower-level convolutions over-extract high-frequency textures at ultra-high resolution, locally weakening the prior constraints of high-level semantics. Future improvements could consider introducing texture suppression modules or superpixel-based region smoothing strategies into shallow features.

[0066] At 2m resolution, FPESNet can still recover the overall outline and shape of medium to large fields relatively well. However, due to the significant smoothing of ground features, strip-shaped irrigation ditches and narrow field ridges gradually become blurred or even indistinguishable in the image. The brightness and texture contrast between fields is weakened, and the boundary representation is relatively blunted. In this case, the model's response to narrow boundaries and small targets is inevitably weakened, manifested as slightly smoothed local outlines and a small number of missed detections in small fields. This is mainly because the deep network completely loses its weak gradient response after multiple downsampling operations. In the future, a feature pyramid structure that maintains high resolution can be used to forcibly preserve the spatial cues of small ground features.

[0067] Overall, FPESNet can stably extract the main field boundaries under different resolution conditions, ensuring the integrity of the shape of the enclosed field and the continuity of the edges. Its performance variation is highly correlated with the objective impact of image resolution on the ability to express ground feature details.

[0068] In high-resolution remote sensing images of farmland, the same field often exhibits complex textures and spectral inconsistencies, as well as spectral similarities between adjacent features. In such complex scenarios, the pure edge detection model DexiNed, which relies solely on local image gradients, is prone to gradient traps, misidentifying harvesting marks or shadows within the field as boundaries and exhibiting breaks at true boundaries with weak spectral contrast. FPESNet introduces a pixel-wise semantic shape auxiliary branch, whose core function is to provide a top-down, object-level prior constraint for edge detection. By establishing pixel-wise semantic representations, the network can mitigate the gradient problems caused by high-frequency details in the feature space.

[0069] Experimental results show that, compared with DexiNed which only outputs edges, FPESNet improves edge IoU by 8.65 percentage points by introducing pixel-wise semantic features as object-level prior knowledge for the edge branch search branch. This further demonstrates that pixel-wise semantics has an optimizing effect on edge features in the field boundary extraction task. However, strong dependence on semantic features also has potential back-interference. When the semantic auxiliary branch causes severe missegmentation in spectrally similar non-farmland areas, this erroneous object-level prior will back-contaminate the edge branch, leading to large areas of false edges or boundary deformation. Therefore, exploring how to dynamically evaluate semantic confidence and adaptively truncate the propagation of erroneous features is a bottleneck that urgently needs to be overcome in this type of dual-branch network.

[0070] FPESNet does not simply concatenate semantic and edge features at the output; instead, it achieves multi-scale fusion through DSCM during the decoding stage. Specifically, each scale combines the semantic branch decoding features and the edge branch features. After residual superposition with the fusion result of the previous layer, it is fed into DSCM. Through the squeeze-excitation and channel attention mechanism, the feature channels that are beneficial to both regional consistency and boundary geometric expression are adaptively highlighted. Then, the morphological features of irregular boundaries in multiple directions and at multiple scales are modeled by combining vertical convolution, horizontal convolution and convolution with different dilation rates.

[0071] Ablation experiments show that adding only MEFM to the baseline improves F1 / IoU to 0.9063 / 0.8287, indicating that multi-scale edge fusion can improve boundary continuity. Further introducing DSCM further improves F1 / IoU to 0.9183 / 0.8490, demonstrating that joint modeling of semantics and edges in the intermediate layer provides additional gain compared to individually enhancing the edge branch. Performance is optimal after stacking FSAAB. Therefore, the proposed fusion strategy effectively injects pixel-wise semantic features into edge features, enabling them to work synergistically in the feature space, thereby enhancing boundary detail expression while maintaining the overall shape consistency of the fields.

[0072] To ensure the model can output the highest quality field edges in complex farmland scenarios and to verify the robustness of multi-task joint optimization, this embodiment modifies the semantic segmentation weights in the joint loss function. ) and edge extraction weights ( A systematic sensitivity analysis was conducted. The core idea of ​​this embodiment is to use semantic segmentation to construct object-level spatial prior knowledge, and then supervise and constrain edge extraction. To explore the optimal collaboration ratio, referring to Table 3, this embodiment sets several representative weight allocation strategies, including 0.5 / 0.5, 0.6 / 0.4, and 0.8 / 0.2.

[0073] When the edge supervision weight is relatively high (e.g.) When edge loss accounts for an excessively large proportion of the total gradient, it forces the model to be extremely sensitive to abrupt gradient changes in local pixels. For example... Figure 14 As shown in (a), due to the lack of sufficiently strong semantic homogeneity priors to suppress internal interference, the feature responses of the DSCM module exhibit large-area, high-intensity scattered activations within the field. This ultimately leads to pseudo-boundaries within the extraction results, such as... Figure 14 (d). However, when semantic weights are too high and edge supervision is insufficient (e.g., ...), In this case, semantic loss dominates the optimization process, causing the model to excessively pursue macro-level regional consistency at the expense of sensitivity to boundary localization. For example... Figure 14 As shown in (c), although high-frequency noise within the fields is completely suppressed, the feature responses at the actual physical boundaries become extremely weak and discontinuous. This over-smoothing causes the boundary lines to blend into the background features, ultimately resulting in blurred segmentation boundaries in the model output, such as... Figure 14 (f). Only when the optimal weight allocation is adopted ( When semantic region constraints and edge spatial localization are balanced, the optimal balance is achieved in feature learning. For example... Figure 14 As shown in (b), a semantic weight of 0.7 forces the network to understand the fields as an indivisible unified whole, filtering out interference from complex internal textures. Meanwhile, as... Figure 14 As shown in (e), the edge weight of 0.3 preserves a moderate and accurate local gradient penalty, resulting in sharp, high-contrast, and continuous feature responses at realistic physical boundaries. With this configuration, the model can simultaneously receive sufficiently strong region integrity gradients and boundary localization gradients, ultimately outputting extraction results with accurate boundaries and complete topological structure.

[0074] Corresponding to the neural network architecture for boundary and morphological representation provided in the above embodiments, this application also provides a system for boundary and morphological representation.

[0075] See Figure 15 The system 20, which is oriented towards boundary and morphological representation, includes: processor 201, memory 202 and communication unit 203.

[0076] These components communicate via one or more buses 204. Those skilled in the art will understand that the system shown in the figure, which is oriented towards boundary and morphological representation, does not constitute a limitation on the embodiments of this application. It can be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0077] The communication unit 203 is used to establish a communication channel, so that the system for boundary and morphological representation can communicate with other devices, such as field image acquisition devices.

[0078] The processor 201 is the control center of the system for boundary and morphological representation. It connects various parts of the system through various interfaces and lines. By running or executing software programs and / or modules stored in memory 202, and calling data stored in memory, it drives the neural network architecture deployed inside to extract field boundaries.

[0079] Memory 202 is used to store the execution instructions of processor 201. Memory 202 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0080] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0081] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A neural network architecture oriented towards boundary and morphological representation, characterized in that, include: The branches include boundary representation, shape semantic modeling, and feature fusion. The boundary representation branch is used to encode the input feature tensor in multiple scale directions based on multiple directional geometric modules to generate a boundary representation tensor. The shape semantic modeling branch is used to perform hierarchical context modeling and structural representation enhancement on the input feature tensor based on multiple convolutional residual units and multiple morphological space-aware adaptive units, so as to generate a region structure representation tensor. The feature fusion branch is used to perform cross-layer interaction, weight modulation, and two-stream combination processing on the boundary representation tensor and the region structure representation tensor to generate the target output feature tensor.

2. The neural network architecture for boundary and morphological representation according to claim 1, characterized in that, The boundary representation branch includes: a preprocessing unit, an directional geometric representation unit, and a multi-scale boundary fusion module (MEFM); wherein: the preprocessing unit is used to perform an initial convolutional transformation on the input feature tensor to generate initial boundary features; the directional geometric representation unit is used to perform stepwise downsampling and directional response enhancement on the initial boundary features to generate boundary feature tensors of multiple scales; and the MEFM is used to perform scale alignment, channel aggregation, and feature reconstruction on the boundary feature tensors of multiple scales to generate boundary representation tensors.

3. The neural network architecture for boundary and morphological representation according to claim 2, characterized in that, The directional geometry representation unit includes multiple directional geometry sub-modules (DGMs). These DGMs decouple and model the directional information in the input features by introducing global convolution, horizontal convolution, and vertical convolution into the network, thereby enhancing the neural network's ability to represent boundary-related responses. The outputs of each branch are aggregated to form the corresponding level's directional enhancement feature tensor. in: Indicates input features, This is a global convolution used to extract the overall structural response. Horizontal convolution is used to extract horizontal response features. Vertical convolution is used to extract vertical response features.

4. The neural network architecture for boundary and morphological representation according to claim 3, characterized in that, The MEFM is used to perform scale alignment, channel aggregation, and feature reconstruction on boundary feature tensors at multiple scales to generate boundary representation tensors, including: After the boundary feature tensors of different scales are input into MEFM, the features at each level are first aligned by convolution, and then uniformly mapped to a common channel space to obtain aligned features. : in: For the first Tensor feature boundaries at various scales It is a linear rectification activation function; After obtaining the alignment features at four scales, respectively... , and Upsampling and alignment to a preset structural scale are performed, and the concatenated data is then combined with convolutional channels before feature reconstruction is performed to complete the fusion of multi-scale boundary representations. in: Indicates upsampling alignment. Indicates to conduct Double upsampling aligns features at each level to the preset structural scale. This is a splicing operation along the channel dimension. This is to compress and splice redundant channels; by As input, after Dropout regularization and After convolution to extract refined features, a multi-channel boundary representation output tensor is generated through Sigmoid activation. : in, express function, Boundary characterization results for each category; Further through Convolution compresses the number of channels to 1, and then normalizes it using a sigmoid function to obtain the final boundary representation output tensor: This is the final output of the boundary representation branch, which will be sent to the subsequent fusion module along with the region structure representation results of the shape semantic modeling branch.

5. The neural network architecture for boundary and morphological representation according to claim 1, characterized in that, The shape semantic modeling branch includes an encoding path, a context aggregation path, and a decoding path. The encoding path consists of multiple basic convolutional residual blocks (BCBlocks) used to extract hierarchical features and maintain gradient propagation stability. The context aggregation path includes a dilated pyramid pooling module (ASPP), which is used to extract and aggregate feature responses under different dilation rates in parallel to enhance the shape semantic modeling branch's ability to represent multi-scale structural relationships and global contextual information. The decoding path is used to perform feature reconstruction step by step and output a region structure representation tensor.

6. The neural network architecture for boundary and morphological representation according to claim 5, characterized in that, The shape semantic modeling branch also includes the morphological space-aware adaptive block FSAAB, which integrates morphological attention and spatial attention mechanisms and achieves the optimal combination of dual-path information through an adaptive weight allocation strategy. The morphological attention is used to model the response intensity of features across different channel dimensions. First, global average pooling and max pooling are performed on the input features to obtain two channel description vectors. These vectors are then processed by two fully connected networks with shared weights and summed. Finally, the vectors are normalized using a sigmoid activation function to form a channel attention map. : The spatial attention is used to model the spatial location distribution of key structural regions. Average pooling and max pooling operations are performed on the input features along the channel dimension, and the concatenated features are then processed through a convolutional layer and a sigmoid activation layer to generate a spatial attention map. : A lightweight weight generator is introduced to dynamically generate two normalized fusion coefficients based on the input global context information, which are used to adjust the strength of channel attention and spatial attention respectively: Finally, the fused output is weighted and corrected using channel attention and spatial attention on the original features: in: It is a channel attention network consisting of two fully connected layers. for, for, This indicates that the channel features are averaged and pooled. This indicates that max pooling is performed on the channel features. This indicates element-wise multiplication. For the adaptive fusion coefficients of the channel attention path, For the adaptive fusion coefficients of the spatial attention path, This is the fusion output feature tensor of FSAAB.

7. The neural network architecture for boundary and morphological representation according to claim 1, characterized in that, The feature fusion branch includes a cross-layer fusion unit and a two-stream combination unit. The cross-layer fusion unit receives decoded features from the shape semantic modeling branch, corresponding scale features from the boundary representation branch, and the fusion result from the previous layer. It then sums or concatenates the decoded features, corresponding scale features, and the fusion result from the previous layer to form a feature tensor to be fused. The two-stream combination unit performs orientation structure reshaping and channel recalibration on the feature tensor to be fused to generate the current layer fused feature tensor.

8. The neural network architecture for boundary and morphological representation according to claim 7, characterized in that, The dual-stream combination unit includes multiple multi-layer dual-stream combination modules (DSCMs). Each DSCM couples integrated features at different scales layer by layer, utilizing a squeeze excitation module (SE) and a channel attention module (SM) to highlight channel features that contribute to the task and suppress redundant responses. Furthermore, vertical and horizontal convolutional channels are introduced to reorganize multi-scale features in the intersecting directions, enabling more comprehensive modeling of the directional dependencies of irregular structures. in: As input features, For extrusion excitation module, For channel attention module, This is a multi-scale feature in the vertical direction obtained by concatenating vertically dilated convolutions at four dilation rates. This is a multi-scale feature in the horizontal direction obtained by concatenating horizontally dilated convolutions at four dilation rates. This is a global oriented multi-scale feature obtained by concatenating global dilated convolutions under four dilation rates. Indicates splicing, , and Represents vertical, horizontal, and global convolution. Indicates the expansion rate; At each layer, the decoded features from the shape semantic modeling branch, the corresponding scale features from the boundary representation branch, and the fusion features generated in the previous layer are additively fused and then input into the DSCM module to enhance the representation of orientation and scale structure. Subsequently, feature reconstruction is performed through transposed convolution, and the number of feature channels is compressed to a smaller dimension to complete the feature reconstruction. Convolution is then used to further smooth and fuse features, suppressing redundant responses introduced during feature reconstruction. Finally, convolution is mapped to the number of categories to obtain the region structure output tensor. in: For the first The fusion output feature tensor of the layers For the first Feature transformation operations of the layered dual-stream combination module Decoding path for shape semantic modeling branch Decoding features of layer output, For the boundary characterization branch Boundary features corresponding to the scale of the layer. For hierarchical restructuring operations, This is a transposed convolution; After obtaining the region structure output tensor, it is jointly encoded with the boundary representation output tensor. Then, the joint representation is reorganized using symmetric feature reconstruction and channel compression operations to generate the final target output feature tensor.

9. The neural network architecture for boundary and morphological representation according to any one of claims 1-8, characterized in that, A region representation loss was constructed. Category equilibrium boundary loss and region-boundary consistency constraint loss Multi-task joint training objective function : in, , and These are the weight coefficients of the corresponding loss function; in: This represents the total number of basic representation units involved in the loss calculation. Indicates the number of categories. For the first The real class label encoding vector corresponding to the basic representation unit is in the _th _ Components on the class, For the first The fundamental representation unit in the th... The predicted output value of the class; in, and These represent the set of real boundary representation units and the set of non-boundary representation units, respectively. The boundary representation value output by the boundary representation branch. and This is a dynamically calculated class balance coefficient used to enhance the constraint effect of sparse boundary samples on parameter updates during backpropagation.

10. The neural network architecture for boundary and morphological representation according to claim 9, characterized in that, By using a differentiable local structure difference operator, difference extraction is performed on the output tensor of the region structure representation to obtain the structural difference representation: In the formula, This represents the convolution operation. Used to truncate outliers after differentiation and normalize them. interval, The output tensor representing the region structure from the shape semantic modeling branch has a differentiable structure difference kernel as follows: After obtaining the structural difference representation, the final consistency constraint term is constructed by calculating the norm distance between it and the output of the boundary representation branch: in: To perform difference extraction on the output tensor representing the region structure using a differentiable local structure difference operator, the th... Structural difference representation values ​​of the locations of the basic representation units For boundary characterization branches in the th The boundary representation output value of each basic representation unit location. Represents norm operations.

Images