High-resolution remote sensing image road automatic extraction method and system based on multi-branch feature guidance and dual-view collaborative decoding

Abstract

The application provides a high-resolution remote sensing image road automatic extraction method and system based on multi-branch feature guidance and dual-view collaborative decoding, and a road segmentation result is obtained through a road extraction network architecture, which comprises a multi-branch feature guidance module, a hierarchical residual encoder, a heterogeneous dual-path decoding architecture composed of a local optimization path and a global optimization path, a path importance self-evaluation module and a final output layer; the multi-branch feature guidance module extracts road primary features with strong robustness through parallel multi-scale perception branches, direction-sensitive branches and context enhancement branches; the hierarchical residual encoder constructs a multi-level feature pyramid to fuse deep semantic features and shallow detail features; the heterogeneous dual-path collaborative decoder comprises a local optimization path and a global optimization path, and realizes pixel-level detail recovery and overall topological connectivity modeling, respectively; the path importance self-evaluation module dynamically generates adaptive fusion weights through a meta-network to realize end-to-end intelligent path collaboration. The application improves the road extraction precision and robustness of high-resolution remote sensing images.

Description

Technical Field

[0001] This invention belongs to the field of photogrammetry and remote sensing image application technology, specifically relating to a method and system for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding. Background Technology

[0002] The continuous development of satellite remote sensing and low-altitude aerial photography technologies has significantly improved the spatial, temporal, and spectral resolution of remote sensing images, providing a richer data foundation for target identification and localization. However, it also places higher technical demands on image interpretation algorithms. Roads, as a core element of basic geographic information systems, are crucial for accurate extraction in fields such as urban planning, traffic management, disaster emergency response, national defense, and socio-economic analysis. Compared to earlier low-resolution images, road targets in high-resolution remote sensing images exhibit more complex morphological features. A single image may contain roads of different grades and materials, with significant differences in their radiometric properties. While providing clear details, high resolution also introduces a large amount of interference information, such as road vehicles, pedestrians, lane markings, and the occlusion and projection of vegetation and buildings on both sides. These factors pose a severe challenge to the accurate and continuous extraction of roads. Therefore, researching road extraction algorithms that can effectively handle complex scenarios and possess both high accuracy and robustness has clear economic and scientific significance.

[0003] Since the 1970s, the field of remote sensing and mapping has systematically explored automatic road extraction technology. Early methods were mainly based on manually designed low-level visual features, constructing rule sets or knowledge-driven models by combining prior geometric features of roads, such as linear continuity, relatively constant width, and radiation characteristics, such as spectral reflectance. These methods showed good interpretability in the context of limited computing resources. However, due to the high variability of road morphology in high-resolution remote sensing images, the complex interference of background features, and the prevalence of occlusion, methods based on handcrafted features have significant limitations in terms of generalization and robustness, as illustrated by patent (CN103996042A). Such methods are not highly adaptable to unstructured roads and complex urban scenes, and are not easy to meet the requirements of completeness and topological accuracy of extraction results in practical engineering applications. With the breakthrough progress of deep learning technology in the field of computer vision, road extraction research ushered in a paradigm shift in the mid-2000s. The introduction of fully convolutional neural networks achieved end-to-end pixel-level prediction for the first time. Through the encoder-decoder architecture, it effectively fused shallow detail features with deep semantic information, significantly improving the accuracy of road boundary localization. These methods automatically learn multi-scale representations of roads through hierarchical stacking of convolutional kernels, avoiding the limitations of manual feature design. However, standard convolutional operations can lead to the loss of spatial details during multiple downsampling processes, especially insufficient restoration of the geometric structure of narrow roads or intersections. On the other hand, the model is highly dependent on large-scale pixel-level labeled data, limiting its cross-regional generalization ability, as shown in patent (CN114882473A). To overcome these limitations, researchers have begun to introduce multi-scale context-aware mechanisms. While these methods can improve the ability to maintain the continuity of main roads, their adaptability to the local geometric deformation of curved roads remains insufficient. Attention mechanisms have also been introduced to enhance the model's ability to select key features and improve the ability to distinguish between roads and non-road areas. However, the introduction of attention modules increases computational complexity and is easily affected by noise in complex backgrounds, as shown in patent (CN117437550A). In addition, graph convolutional networks further enhance the ability to maintain topological integrity by explicitly modeling the spatial relationships between road pixels, but the construction quality of the graph structure is highly dependent on the node sampling strategy, making it sensitive to noise and computationally inefficient. In addition, generative adversarial networks are used to improve the structural realism of the extracted results and constrain the rationality of the road topology of the generated results through the discriminator network. However, the training process is unstable and prone to mode collapse.

[0004] Existing road extraction methods based on multi-scale feature fusion, such as patent CN120411784A, enhance the scale adaptability of road features through multi-scale and multi-directional feature fusion modules. However, their feature fusion mechanisms remain at the level of simple feature stitching or weighted summation, failing to achieve truly adaptive feature selection. This fixed-weight fusion method struggles to cope with the high variability of road morphology in high-resolution images, especially in curved road sections and intersections, where insufficient or over-response of features is prone to occur. Furthermore, while the shallow feature-guided attention module used in this scheme can improve detail reconstruction capabilities, its attention weight generation process lacks explicit modeling of the global topology. In terms of structural design, this scheme is based on the D-LinkNet decoding path architecture. Although it strengthens direction awareness through multi-directional feature enhancement modules, the two mutually constraining goals of local detail restoration and global topology understanding are forcibly compressed into the same path, making it difficult for the network to balance pixel-level accuracy and topological connectivity in complex scenes. Despite significant progress on specific datasets, a deeper analysis of its technical implementation and performance reveals several core bottlenecks hindering its large-scale application in real-world surveying and mapping scenarios. These limitations primarily manifest in key dimensions such as the inherent contradictions in the multi-scale feature fusion mechanism, the imbalance between local details and global topology understanding, and insufficient protection of topological integrity in complex scenarios.

[0005] To address the aforementioned technical bottlenecks, this invention aims to construct a new paradigm for road extraction that can simultaneously preserve local details and provide global topological understanding. It focuses on overcoming the performance limitations of existing single-path models by organically integrating multi-branch feature guidance and heterogeneous dual-path collaborative decoding to achieve end-to-end road extraction optimization. Summary of the Invention

[0006] The purpose of this invention is to propose an automatic road extraction method and system for high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding. It aims to overcome the technical bottlenecks of existing road extraction technologies in feature representation, multi-scale fusion and topology modeling. Through innovative network architecture design, it achieves a significant improvement in the accuracy and robustness of road extraction from high-resolution remote sensing images, and promotes the implementation of this technology in practical surveying and mapping applications.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] The automatic road extraction method for high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding includes the following steps:

[0009] Acquire high-resolution remote sensing images;

[0010] High-resolution remote sensing images are input into a pre-constructed end-to-end road extraction network architecture, which includes a multi-branch feature guidance module (MFGM), a hierarchical residual encoder, a heterogeneous dual-path collaborative decoder (HDCD) based on the collaboration of local and global optimization paths, and a path importance self-evaluation module (PISM).

[0011] The input image is subjected to front-end feature extraction by the branch feature guidance module, and the enhanced primary feature representation is output. The primary feature representation is input to the hierarchical residual encoder for deep feature extraction to generate a multi-scale feature pyramid. The multi-scale feature pyramid is input to the heterogeneous dual-path collaborative decoder composed of local optimization path and global optimization path for feature reconstruction and road segmentation. The output features of the two heterogeneous paths are adaptively fused by the path importance self-evaluation module, and the road segmentation result is obtained based on the final fused features.

[0012] Preferably, the multi-branch feature guidance module (MFGM) includes a multi-scale perception branch, a direction-sensitive branch, and a context-enhancing branch set in parallel. The output features of the three branches are concatenated in the channel dimension, adaptively weighted by the channel attention fusion module, and refined by the refinement network to obtain an enhanced primary feature representation.

[0013] Preferably, the multi-scale sensing branch, orientation-sensitive branch, context-enhancing branch, and refined network are specifically as follows:

[0014] The multi-scale sensing branch performs 3×3, 5×5 and 7×7 convolutions on the input feature map respectively, and concatenates the three convolution outputs into the first-level feature to obtain the multi-scale sensing features.

[0015] The orientation-sensitive branch performs 1×7 and 1×3 horizontal bar convolutions and 7×1 and 3×1 vertical bar convolutions on the input feature map to extract horizontal and vertical features. After rotating the input feature map by 45° and 135°, it reuses the 7×1 and 3×1 vertical bar convolution kernels for convolution operations, and then rotates the convolution results by -45° and -135° respectively to return to the correct orientation, thereby extracting diagonal road features in the 45° and 135° directions. The horizontal, vertical, 45°, and 135° diagonal road features are then concatenated to obtain the orientation-sensitive features.

[0016] The context enhancement branch obtains context enhancement features by performing cascaded dilated convolutions with dilation rates of 2 and 4 on the input feature map;

[0017] The refined network consists of 1×1 convolutions and 3×3 convolutions.

[0018] Preferably, the hierarchical residual encoder adopts a four-stage hierarchical structure:

[0019] In the first stage, the feature map output by MFGM is processed through three residual blocks, adjusted by a multi-scale fusion gate, and then fused with shallow detail features obtained from the original input image through downsampling and convolution transformation via skip connections. The output is a first-level feature map with a spatial size of half that of the original input.

[0020] The second stage processes the first-level feature map through four residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through skip connections. The output is a second-level feature map with a spatial size that is one-quarter of the original input.

[0021] The third stage processes the second-level feature map through 6 residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through deeper downsampling and convolution transformation through skip connections. The output is a third-level feature map with a spatial size of one-eighth of the original input.

[0022] The fourth stage processes the third-level feature map through three residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through the deepest level downsampling and convolution transformation through skip connections. The output is a fourth-level feature map with a spatial size of one-sixteenth of the original input.

[0023] Each residual block contains two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, and is equipped with a shortcut connection to pass the input features to the output of the residual block and fuse them with the output feature map of the residual branch.

[0024] Preferably, the heterogeneous dual-path collaborative decoder includes a parallel local optimization path, a global optimization path, and a set of multi-scale fusion gates; the multi-scale fusion gates are used to adjust the scale and channels of the output features of each level of the encoder before skip connections.

[0025] The local optimization path first upsamples the input features of the current decoding level by 2 times through a 4×4 transposed convolution at each decoding stage. At the same time, it constructs an independent residual connection path, first upsamples the input features to the same size, and then adjusts the channels through a 1×1 convolution and batch normalization. Finally, the outputs of the two paths are added and passed through an activation function to obtain the output features of the current decoding level. Then, the corresponding level features of the encoder adjusted by the multi-scale fusion gate are fused to obtain the input features of the next decoding level.

[0026] The global optimization path integrates a multi-scale dilated convolutional pyramid structure. It extracts multi-scale contextual information from the input features of the current decoding level by using 3×3 dilated convolutional layers with dilation rates of 1, 2, 4, and 8 in parallel. After concatenating and fusing the outputs of each dilated convolution, it upsamples the output features of the current decoding level through transposed convolution to obtain the output features of the current decoding level. Then, it fuses the corresponding level features of the encoder adjusted by the multi-scale fusion gate to obtain the input features of the next decoding level.

[0027] Preferably, the adaptive fusion process of the path importance self-evaluation module includes:

[0028] Global average pooling is performed on the local optimization path output feature Flocal and the global optimization path output feature Fglobal respectively to obtain two C-dimensional global feature vectors glocal and gglobal;

[0029] Two global feature vectors, glocal and gglobal, are concatenated to form a 2C-dimensional joint representation, which is then input into a miniature fully connected neural network. The miniature fully connected neural network includes a 2C-dimensional input layer, a 32-dimensional hidden layer using the ReLU activation function, and a 2-dimensional output layer.

[0030] The output features of the output layer are processed by the Softmax function to generate local path weights wlocal and global path weights wglobal, and the local path weights and global path weights satisfy wlocal + wglobal = 1.

[0031] The local optimized path output features and the global optimized path output features are weighted and summed pixel by pixel based on the local path weights and global path weights to obtain the final fused feature Ffinal = wlocal·Flocal + wglobal·Fglobal.

[0032] The automatic road extraction system for high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the above-mentioned automatic road extraction method for high-resolution remote sensing images.

[0033] Preferably, the system also integrates a human-computer interaction interface, supporting users to make visual adjustments and interactive optimizations to the extraction results.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] This invention, through innovative technologies such as multi-branch feature guidance, heterogeneous dual-path collaborative decoding, and adaptive fusion, demonstrates significantly superior performance compared to existing technologies in terms of feature representation capabilities, detail preservation and topology understanding synergy, adaptive fusion, computational efficiency, robustness to complex scenes, and end-to-end processing integrity. These technological advantages not only improve the accuracy and reliability of road extraction but also provide new technical ideas and methodological support for the intelligent interpretation of high-resolution remote sensing images. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a structural diagram of the multi-branch feature guidance module of the present invention;

[0038] Figure 3 This is a structural diagram of the hierarchical residual encoder (part a in the figure) and the heterogeneous dual-channel collaborative decoder (parts b and c in the figure) of the present invention;

[0039] Figure 4 This is a structural diagram of the path importance self-assessment module of the present invention;

[0040] Figure 5 This is a schematic diagram of the human-computer interaction interface of the present invention. Detailed Implementation

[0041] The following is in conjunction with the appendix Figure 1-5 The technical solution of the present invention will be described in detail below.

[0042] This invention proposes an automatic road extraction method and system for high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding to address the following key technical issues: First, a feature extraction mechanism capable of simultaneously capturing multi-scale context, orientation-sensitive features, and long-range dependencies is designed to enhance the model's ability to represent complex road structures. Second, a collaborative decoding framework integrating local detail optimization and global topological understanding is established to resolve the inherent contradiction between detail preservation and connectivity assurance in a single decoding path. Third, an adaptive feature fusion strategy is developed to dynamically adjust the contribution weights of different paths based on the input image content, achieving intelligent feature selection and integration. Fourth, the model's computational efficiency is optimized while ensuring extraction accuracy, enabling it to meet the real-time processing requirements of high-resolution remote sensing images. This invention systematically addresses the above technical challenges through the collaborative design of a multi-branch feature guidance module, a heterogeneous dual-path collaborative decoder, and a path importance self-evaluation module. The multi-branch feature guidance module integrates multi-scale perception, orientation-sensitive convolution, and context enhancement functions through a parallel architecture, providing a robust feature foundation for subsequent processing. The heterogeneous dual-path collaborative decoder employs a functionally specialized dual-path design, focusing on pixel-level detail recovery and overall topology modeling respectively. It also achieves content-aware dynamic feature fusion through a path importance self-evaluation module. Experimental verification on the DeepGlobe and RoadTracer datasets demonstrates that the above-mentioned technical approach significantly improves the completeness and topological accuracy of road extraction, particularly showing a clear advantage in restoring the connectivity of broken roads and reconstructing the completeness of complex intersection structures.

[0043] This invention proposes an automatic road extraction method and system for high-resolution remote sensing images based on multi-branch feature guidance and heterogeneous dual-path collaborative decoding. It is applicable to automatic road information extraction tasks from high-resolution remote sensing images in multiple fields such as urban planning, traffic management, and disaster emergency response. The core technical solution of this method constructs an end-to-end road extraction network architecture through the collaborative design of a multi-branch feature guidance module (MFGM), a hierarchical residual encoder, a heterogeneous dual-path collaborative decoder (HDCD), and a path importance self-evaluation module (PISM). The processing flow begins with the input high-resolution remote sensing image, undergoes multi-stage feature extraction and fusion, and finally outputs accurate road segmentation results. The multi-branch feature guidance module, as the network front-end, provides enhanced feature representations for subsequent processing; the hierarchical residual encoder constructs a multi-level feature pyramid; the heterogeneous dual-path decoder optimizes local details and global topology respectively; and the path importance self-evaluation module achieves adaptive fusion of the dual-path outputs. This modular design not only ensures the independence of each functional unit but also achieves a synergistic improvement in overall performance through carefully designed interfaces.

[0044] The input image first enters the Multi-Branch Feature Guidance Module (MFGM), which serves as the network's front-end feature extractor. MFGM employs a parallel three-branch architecture, responsible for multi-scale perception, orientation-sensitive feature extraction, and context enhancement, respectively. The multi-scale perception branch uses convolutional kernels of three different sizes (3×3, 5×5, and 7×7) in parallel to simultaneously capture local details and global structural features of the road, fundamentally improving the network's ability to perceive road features. The orientation-sensitive branch, targeting the linear geometric characteristics of roads, uses 1×7 and 1×3 strip convolutional kernels to extract horizontal features and 7×1 and 3×1 convolutional kernels to extract vertical features, and achieves adaptive enhancement for diagonal roads through rotation operations. The context enhancement branch expands the receptive field without reducing feature map resolution by cascading dilated convolutional layers with dilation rates of 2 and 4, capturing dependencies between distant pixels. The output features of the three branches are concatenated along the channel dimension, adaptively weighted by the channel attention fusion module, and finally refined through a convolutional network consisting of 1×1 and 3×3 convolutions to output an enhanced primary feature representation. This multi-dimensional feature guidance mechanism effectively overcomes the limitations of traditional single-path feature extraction and provides a more robust primary feature representation for subsequent processing.

[0045] The features preprocessed by the MFGM then enter a hierarchical residual encoder for deep feature extraction. The encoder employs a four-stage hierarchical structure, with each stage built around a residual learning unit. The first stage receives the feature map output from the MFGM and transforms it through three residual blocks. Each residual block contains two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, equipped with shortcut connections to ensure effective gradient propagation. The spatial size of the output feature map in this stage is reduced to half that of the original input. The second stage further deepens the feature representation through four residual blocks, introducing convolutions with a stride of 2 for downsampling, reducing the feature map size to one-quarter of the input. The third stage, the core component of the encoder, contains six residual blocks, aiming to capture more discriminative road structure patterns in the middle layers. The output feature size is one-eighth of the input. The final stage extracts global contextual information with the widest receptive field through three residual blocks, reducing the feature map size to one-sixteenth of the input. Each encoding stage integrates a multi-scale fusion gate mechanism, fusing detailed features from shallow layers through skip connections to ensure effective preservation of deep semantic information and shallow spatial details.

[0046] The multi-scale feature pyramid generated by the encoder is then fed into a heterogeneous dual-path collaborative decoder (HDCD) for feature reconstruction and road segmentation. HDCD abandons the traditional single decoding path, employing a specialized dual-path design, including a Local Optimization Path (LRP) and a Global Optimization Path (GCP). The LRP focuses on pixel-level detail recovery, using a modified residual connection mechanism for efficient upsampling. In each decoding stage, this path first performs a 2x upsampling through a 4×4 transposed convolution, followed by channel adjustment of the input features through independent 1×1 convolutions, and then sums the two to maximize information flow. Simultaneously, the LRP fuses detailed, high-resolution features from the corresponding encoder layer through skip connections, providing a reference for accurate reconstruction of local geometry. The GCP focuses on topological connectivity modeling of the road network, integrating a multi-scale dilated convolutional pyramid structure. By using 3×3 dilated convolutional layers with dilation rates of 1, 2, 4, and 8 in parallel, it rapidly expands the receptive field while maintaining the spatial resolution of the feature map, capturing multi-scale contextual information from local to global perspectives. After the outputs of each branch are concatenated and fused, they are also upsampled through transposed convolution and combined with encoder features to enhance global semantic understanding. The two paths process features from the previous decoding layer and encoder features introduced through skip connections at various stages of the decoding process, but their focuses are distinctly different. This functional specialization and complementarity ensures that the final output is both rich in detail and structurally coherent. Through this functionally specialized dual-path collaborative mechanism, the topological integrity of the road network is significantly improved while maintaining high positioning accuracy, exhibiting excellent resistance to occlusion and complex background interference.

[0047] The output features of the two heterogeneous paths are adaptively fused by the Path Importance Self-Evaluation (PISM) module. A meta-learning network dynamically evaluates and assigns weights to the outputs of the two decoded paths in the final fusion. First, global average pooling is performed on the local path output feature Flocal and the global path output feature Fglobal, respectively, resulting in two C-dimensional global feature vectors glocal and gglobal. These two vectors are concatenated to form a 2C-dimensional joint representation, which is then fed into a miniature fully connected neural network for analysis. This network consists of an input layer (2C dimension), hidden layers (32 dimensions, using the ReLU activation function), and an output layer (2 dimensions). Finally, the local path weight wlocal and the global path weight wglobal are generated using the Softmax function, ensuring that their sum is 1. The fusion process is represented as a pixel-wise weighted summation: Ffinal = wlocal·Flocal + wglobal·Fglobal. This content-aware dynamic weight allocation mechanism enables the network to automatically adjust the contribution of the two paths according to the characteristics of the input image, achieving intelligent feature collaboration and effectively overcoming the limitations of fixed-weight fusion strategies. The features, after adaptive fusion by the path importance self-evaluation module, are fed into the final output layer. Through an operation sequence containing 3×3 convolutions and 1×1 convolutions, the number of feature channels is mapped to the preset number of road categories, generating the final road segmentation probability map.

[0048] This technical solution constructs a complete road extraction workflow through the organic integration of the aforementioned modules. The entire network uses the front-end MFGM module to achieve multi-scale, multi-directional road feature enhancement extraction; a hierarchical residual encoder is used to construct a deep-fusion feature pyramid; heterogeneous dual-path decoders are used to optimize local details and global topology respectively; and finally, the PISM module is used to achieve adaptive fusion of dual-path outputs. Experimental validation on the DeepGlobe and RoadTracer datasets shows that this method can significantly improve the completeness and topological accuracy of road extraction, especially demonstrating significant advantages in restoring the connectivity of broken roads and the complete restoration of complex intersection structures. The system also integrates a human-computer interaction interface, supporting users to visually adjust and interactively optimize the extraction results, further improving the convenience and reliability of practical applications.

[0049] The high-resolution remote sensing image automatic road extraction system proposed in this invention, based on multi-branch feature guidance and heterogeneous dual-path collaborative decoding, also incorporates an innovative human-computer interaction interface design, supporting users to visually adjust and interactively optimize the extraction results. This system not only implements an end-to-end road extraction process but also provides real-time result feedback, allowing users to directly edit and correct the extraction results through the interface, greatly improving operational efficiency and user experience. This human-computer collaborative working mode leverages the efficiency advantages of automated computer processing while incorporating the semantic understanding capabilities of human experts, making it of significant value in practical applications.

[0050] In summary, this invention significantly improves the topological integrity of the road network while maintaining high positioning accuracy, and has excellent resistance to occlusion and complex background interference.

[0051] First, this invention demonstrates a significant advantage in representing road features. Existing methods based on single-scale or simple multi-scale fusion often struggle to simultaneously capture both local details and global contextual information. This invention effectively addresses this limitation by designing a multi-branch feature guidance module. This module employs a parallel three-branch architecture. The multi-scale perception branch, through the collaborative work of three different convolutional kernel sizes (3×3, 5×5, and 7×7), achieves cross-scale feature capture from local texture to global structure. The orientation-sensitive branch, targeting the linear geometric characteristics of roads, enhances the perception of roads with different orientations through a combination of 1×7 and 1×3 strip convolutional kernels and rotation operations. The context enhancement branch, through cascaded dilated convolutional layers with dilation rates of 2 and 4, significantly expands the receptive field without sacrificing spatial resolution. This multi-dimensional feature guidance mechanism enables the network to establish a comprehensive understanding of the road structure at the initial stage of feature extraction, laying a solid foundation for subsequent processing. Experimental results demonstrate the superiority of this design in feature representation on the DeepGlobe dataset.

[0052] Secondly, this invention achieves significant progress in the collaborative optimization of detail preservation and topology understanding. Most existing methods employ a single decoding path, making it difficult to simultaneously meet the dual requirements of pixel-level accuracy and topological integrity. This invention innovatively proposes a heterogeneous dual-path collaborative decoder architecture, where the local optimization path focuses on achieving efficient upsampling through a modified residual connection mechanism, ensuring accurate recovery of road edge details; the global optimization path strengthens the modeling of long-distance dependencies through a multi-scale dilated convolutional pyramid structure, guaranteeing the connectivity of the road network. The specialized design and collaborative working mechanism of the two paths fundamentally resolve the inherent contradiction between detail preservation and topology understanding in traditional methods. Figure 1The overall framework of the present invention clearly demonstrates this dual-path collaborative working mechanism, in which the local optimization path and the global optimization path process feature information at different levels respectively, and finally achieve complementary advantages through intelligent fusion.

[0053] Third, this invention achieves a technological breakthrough in the adaptability of feature fusion. Existing feature fusion technologies often employ fixed weights or simple weighting methods, making it difficult to adapt to changes in feature distribution across different scenarios. The path importance self-evaluation module proposed in this invention dynamically generates adaptive fusion weights through a meta-learning network, automatically adjusting the contribution of the two decoding paths based on the content characteristics of the input image. This module first performs global average pooling on the output features of the local and global paths respectively, obtaining their respective global statistical feature vectors. Then, it analyzes the joint representation of these vectors through a mini fully connected neural network, and finally generates normalized fusion weights through a Softmax function. This content-aware dynamic fusion mechanism enables the network to flexibly cope with various complex scenarios, significantly improving the model's generalization ability. Experimental results on the RoadTracer dataset demonstrate its effectiveness in practical applications.

[0054] Fourth, this invention demonstrates significant advantages in robustness under complex scenarios. Road extraction from high-resolution remote sensing imagery faces multiple challenges, including occlusion, shadows, and varied morphologies, and existing methods often perform poorly in these complex scenarios. This invention effectively enhances the model's anti-interference capability through multi-layered technological innovations: the multi-branch feature guidance module enhances the model's robustness to noise and background interference; the heterogeneous dual-path collaborative decoder ensures the preservation of topological continuity under local occlusion; and the adaptive fusion mechanism enables the model to dynamically adjust its processing strategy according to scene characteristics. Qualitative comparison results on public datasets clearly demonstrate the superior performance of this method in complex scenarios, particularly its outstanding ability to restore the connectivity of broken roads and the structural integrity of complex intersections.

[0055] Finally, this invention achieves significant improvements in the completeness of the end-to-end processing flow. Through the collaborative work of its internal modules, this invention realizes a complete processing flow from raw image input to the final road segmentation result. The multi-branch feature guidance module is responsible for primary feature enhancement, the hierarchical residual encoder constructs deep feature representations, the heterogeneous dual-path collaborative decoder achieves collaborative optimization of details and topology, and the path importance self-evaluation module completes intelligent fusion decision-making. This complete processing chain ensures the accuracy and reliability of the final result. This end-to-end design not only simplifies the application process but also enhances the practical value of the system.

[0056] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding, characterized in that, Specifically, the following steps are included: Acquire high-resolution remote sensing images; High-resolution remote sensing images are input into a pre-constructed end-to-end road extraction network architecture, which includes a multi-branch feature guidance module (MFGM), a hierarchical residual encoder, a heterogeneous dual-path collaborative decoder (HDCD) based on the collaboration of local and global optimization paths, and a path importance self-evaluation module (PISM). The input image is subjected to front-end feature extraction by the branch feature guidance module, and the enhanced primary feature representation is output. The primary feature representation is input to the hierarchical residual encoder for deep feature extraction to generate a multi-scale feature pyramid. The multi-scale feature pyramid is input to the heterogeneous dual-path collaborative decoder composed of local optimization path and global optimization path for feature reconstruction and road segmentation. The output features of the two heterogeneous paths are adaptively fused by the path importance self-evaluation module, and the road segmentation result is obtained based on the final fused features.

2. The method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding as described in claim 1, characterized in that, The multi-branch feature guidance module (MFGM) includes a multi-scale perception branch, a direction-sensitive branch, and a context-enhancing branch set in parallel. The output features of the three branches are concatenated in the channel dimension, adaptively weighted by the channel attention fusion module, and then refined by the refinement network to obtain the enhanced primary feature representation.

3. The method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding according to claim 2, characterized in that, The multi-scale sensing branch, orientation-sensitive branch, context-enhancing branch, and refined network are detailed below: The multi-scale sensing branch performs 3×3, 5×5 and 7×7 convolutions on the input feature map respectively, and concatenates the three convolution outputs into the first-level feature to obtain the multi-scale sensing features. The orientation-sensitive branch performs 1×7 and 1×3 horizontal bar convolutions and 7×1 and 3×1 vertical bar convolutions on the input feature map to extract horizontal and vertical features. After rotating the input feature map by 45° and 135°, it reuses the 7×1 and 3×1 vertical bar convolution kernels for convolution operations, and then rotates the convolution results by -45° and -135° respectively to return to the correct orientation, thereby extracting diagonal road features in the 45° and 135° directions. The horizontal, vertical, 45°, and 135° diagonal road features are then concatenated to obtain the orientation-sensitive features. The context enhancement branch obtains context enhancement features by performing cascaded dilated convolutions with dilation rates of 2 and 4 on the input feature map; The refined network consists of 1×1 convolutions and 3×3 convolutions.

4. The method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding according to claim 3, characterized in that, The hierarchical residual encoder adopts a four-stage hierarchical structure: In the first stage, the feature map output by MFGM is processed through three residual blocks, adjusted by a multi-scale fusion gate, and then fused with shallow detail features obtained from the original input image through downsampling and convolution transformation via skip connections. The output is a first-level feature map with a spatial size of half that of the original input. The second stage processes the first-level feature map through four residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through skip connections. The output is a second-level feature map with a spatial size that is one-quarter of the original input. The third stage processes the second-level feature map through 6 residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through deeper downsampling and convolution transformation through skip connections. The output is a third-level feature map with a spatial size of one-eighth of the original input. The fourth stage processes the third-level feature map through three residual blocks, adjusts it through a multi-scale fusion gate, and then fuses the shallow detail features obtained from the original input image through the deepest level downsampling and convolution transformation through skip connections. The output is a fourth-level feature map with a spatial size of one-sixteenth of the original input. Each residual block contains two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, and is equipped with a shortcut connection to pass the input features to the output of the residual block and fuse them with the output feature map of the residual branch.

5. The method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding according to claim 4, characterized in that, The heterogeneous dual-path collaborative decoder includes parallel local optimization paths, global optimization paths, and a set of multi-scale fusion gates for collaborative implementation; the multi-scale fusion gates are used to adjust the scale and channels of the output features of each level of the encoder before skip connections; The local optimization path first upsamples the input features of the current decoding level by 2 times through a 4×4 transposed convolution at each decoding stage. At the same time, it constructs an independent residual connection path, first upsamples the input features to the same size, and then adjusts the channels through a 1×1 convolution and batch normalization. Finally, the outputs of the two paths are added and passed through an activation function to obtain the output features of the current decoding level. Then, the corresponding level features of the encoder adjusted by the multi-scale fusion gate are fused to obtain the input features of the next decoding level. The global optimization path integrates a multi-scale dilated convolutional pyramid structure. It extracts multi-scale contextual information from the input features of the current decoding level by using 3×3 dilated convolutional layers with dilation rates of 1, 2, 4, and 8 in parallel. After concatenating and fusing the outputs of each dilated convolution, it upsamples the output features of the current decoding level through transposed convolution to obtain the output features of the current decoding level. Then, it fuses the corresponding level features of the encoder adjusted by the multi-scale fusion gate to obtain the input features of the next decoding level.

6. The method for automatic road extraction from high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding according to claim 5, characterized in that, The adaptive fusion process of the path importance self-evaluation module includes: Global average pooling is performed on the local optimization path output feature Flocal and the global optimization path output feature Fglobal respectively to obtain two C-dimensional global feature vectors glocal and gglobal; Two global feature vectors, glocal and gglobal, are concatenated to form a 2C-dimensional joint representation, which is then input into a miniature fully connected neural network. The miniature fully connected neural network includes a 2C-dimensional input layer, a 32-dimensional hidden layer using the ReLU activation function, and a 2-dimensional output layer. The output features of the output layer are processed by the Softmax function to generate local path weights wlocal and global path weights wglobal, and the local path weights and global path weights satisfy wlocal + wglobal = 1. The local optimized path output features and the global optimized path output features are weighted and summed pixel by pixel based on the local path weights and global path weights to obtain the final fused feature Ffinal = wlocal·Flocal + wglobal·Fglobal.

7. A high-resolution remote sensing image automatic road extraction system based on multi-branch feature guidance and dual-view collaborative decoding, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the automatic road extraction method for high-resolution remote sensing images as described in any one of claims 1-6.

8. The automatic road extraction system for high-resolution remote sensing images based on multi-branch feature guidance and dual-view collaborative decoding according to claim 7, characterized in that, The system also integrates a human-computer interaction interface, allowing users to visually adjust and optimize the extraction results.

Images