Road blockage detection methods, devices, electronic equipment and storage media

Abstract

This invention provides a road blockage detection method, apparatus, electronic device, and storage medium, relating to the field of image processing technology. The method includes: applying edge structure enhancement to pre-disaster images using a temporal-specific enhancement module to highlight the linear features of slender roads; applying spatial region enhancement to post-disaster images to strengthen the blocky consistency of landslides and collapses; and extracting prior information on dual-temporal differences, effectively solving the feature mismatch problem caused by significant differences in target morphology between pre- and post-disaster images. A bidirectional cross-guided fusion module achieves mutual constraint and collaborative fusion of local detail features and global semantic features, avoiding feature fragmentation caused by simple fusion. A spatial gating enhancement module uses temporal difference priors to generate gating weights for non-destructive additive enhancement of the fused features, suppressing background noise while improving the response intensity of low-proportion change areas. The synergistic effect of these three components improves the accuracy of road blockage detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a road obstruction detection method, apparatus, electronic device, and storage medium. Background Technology

[0002] Natural disasters often cause severe damage to road networks, leading to traffic disruptions and seriously affecting emergency rescue and material transportation. Quickly and accurately identifying road blockage areas is crucial for disaster emergency response.

[0003] High-resolution satellite imagery can acquire extensive surface information within hours after a disaster, and road blockage areas can be identified by comparing pre- and post-disaster images. However, existing deep learning-based change detection methods face the following technical challenges in road blockage detection: pre-disaster roads have a slender linear structure, while post-disaster landslides are irregular clumps. Existing methods use the same feature extraction strategy for both temporal images, making it difficult to adapt to the representation needs of the two types of targets; convolutional neural networks have limited receptive fields, and the Transformer's local details are blurred, making simple fusion prone to feature fragmentation; the blockage area has a small pixel ratio, and end-to-end segmentation loss alone is insufficient for supervision, resulting in slow model convergence and weak generalization.

[0004] Therefore, how to accurately detect road blockage information has become an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a road blockage detection method, apparatus, electronic device, and storage medium to address the shortcomings of low accuracy in existing road blockage detection technologies and improve the accuracy of road blockage detection.

[0006] This invention provides a road obstruction detection method, comprising the following steps: Acquire pre-disaster and post-disaster images to be inspected; The pre-disaster and post-disaster images are input into a target detection model to obtain the change detection results of road blockage areas output by the target detection model; wherein, the target detection model includes an inter-guide neck network, and the inter-guide neck network includes: The temporal-specific enhancement module is used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference priors based on the differences between the pre-disaster local features and the post-disaster local features. The bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster images and the post-disaster images with the temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference prior. The spatial gating enhancement module is used to enhance the cross-guidance fusion features based on the enhanced pre-disaster local features, the enhanced post-disaster local features, and the spatial gating map generated by the temporal difference prior, and output the enhanced fusion features; the change detection result of the road blockage area is generated based on the enhanced fusion features.

[0007] According to a road obstruction detection method provided by the present invention, the temporal-phase specific enhancement module is specifically used for: The enhanced pre-disaster local features are obtained by performing an edge attention operation on the pre-disaster local features. The enhanced post-disaster local features are obtained by performing a spatial attention operation on the post-disaster local features. The difference between the pre-disaster local features and the post-disaster local features is calculated pixel by pixel to obtain the difference features. The difference features are then transformed through convolution to obtain the temporal difference prior.

[0008] According to a road obstruction detection method provided by the present invention, the bidirectional intersection guidance fusion module is specifically used for: The shared global semantic features are used to generate a first spatial guiding weight for the temporal-specific fusion features, and the first spatial guiding weight is used to weight the temporal-specific fusion features. The temporal-specific fusion features are used to generate a second spatial guiding weight for the shared global semantic features, and the shared global semantic features are weighted using the second spatial guiding weight; The weighted temporal-specific fusion features are fused with the weighted shared global semantic features to output the cross-guided fusion features.

[0009] According to a road obstruction detection method provided by the present invention, the spatial gating enhancement module is specifically used for: The enhanced pre-disaster local features, the enhanced post-disaster local features, and the temporal difference prior are spliced ​​together to obtain spliced ​​features; The stitching features are processed by a gating predictor and activated by an activation function to generate a spatial gating map; Using the spatial gating graph, the cross-guided fusion feature is non-destructively additively enhanced to obtain the enhanced fusion feature.

[0010] According to a road obstruction detection method provided by the present invention, the target detection model further includes a dual-branch feature extraction backbone network, wherein the dual-branch feature extraction backbone network includes: The first feature extraction branch is used to extract the pre-disaster local features of the pre-disaster image and the post-disaster local features of the post-disaster image. The second feature extraction branch is used to extract the shared global semantic features of the pre-disaster images and the post-disaster images.

[0011] According to a road obstruction detection method provided by the present invention, the target detection model further includes a decoder, which is used to perform multi-scale context aggregation on the enhanced fusion features to generate a change detection result of the road obstruction area.

[0012] According to a road obstruction detection method provided by the present invention, the decoder is specifically used for: The enhanced fusion features are multi-scale pooled by the pyramid pooling module to obtain multi-scale pooled features. The multi-scale pooling features and the enhanced fusion features of each layer are fused from top to bottom using a feature pyramid network to obtain multi-scale fusion features. Based on the multi-scale fusion features, a change detection result for the road obstruction area is generated. This invention also provides a road obstruction detection device, comprising the following modules: The acquisition module is used to acquire pre-disaster and post-disaster images to be detected. A road blockage detection module is used to input the pre-disaster image and the post-disaster image into a target detection model, and obtain the change detection results of the road blockage area output by the target detection model; wherein, the target detection model includes an inter-guide neck network, and the inter-guide neck network includes: The temporal-specific enhancement module is used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference priors based on the differences between the pre-disaster local features and the post-disaster local features. The bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster images and the post-disaster images with the temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference prior. The spatial gating enhancement module is used to enhance the cross-guidance fusion features based on the enhanced pre-disaster local features, the enhanced post-disaster local features, and the spatial gating map generated by the temporal difference prior, and output the enhanced fusion features; the change detection result of the road blockage area is generated based on the enhanced fusion features.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the road obstruction detection method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the road obstruction detection method as described above.

[0015] The road blockage detection method, apparatus, electronic device, and storage medium provided by this invention employ a temporal-specific enhancement module to apply edge structure enhancement to pre-disaster images to highlight the linear features of slender roads, and a spatial region enhancement module to apply spatial region enhancement to post-disaster images to strengthen the blocky consistency of landslides and collapses. It also extracts prior information on dual-temporal differences, effectively solving the feature mismatch problem caused by significant differences in target morphology between pre- and post-disaster images. A bidirectional cross-guided fusion module achieves mutual constraint and collaborative fusion of local detailed features and global semantic features, avoiding feature fragmentation caused by simple fusion. A spatial gating enhancement module uses temporal difference priors to generate gating weights for non-destructive additive enhancement of the fused features, suppressing background noise while increasing the response intensity of low-proportion change areas. The synergistic effect of these three components improves the accuracy of road blockage detection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of the road obstruction detection method provided by the present invention.

[0018] Figure 2 This is the overall architecture diagram of RoadGuard-Net provided by this invention.

[0019] Figure 3 This is the flowchart of RoadGuard-Net provided by the present invention.

[0020] Figure 4 This is a schematic diagram of the key sub-modules of TSM provided by the present invention.

[0021] Figure 5 This is a schematic diagram of the CrossGuidedNeck structure provided by the present invention.

[0022] Figure 6This is a schematic diagram of the UPerHead decoding structure provided by the present invention.

[0023] Figure 7 This is a visualization comparison chart of the BlockedRoad dataset provided by this invention.

[0024] Figure 8 This is a visualization comparison chart of the LEVIR-CD dataset provided by this invention.

[0025] Figure 9 This is a schematic diagram of the road blockage detection device provided by the present invention.

[0026] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] The following is combined with Figures 1-10 The present invention describes a road obstruction detection method, apparatus, electronic device, and storage medium.

[0029] Figure 1 This is a schematic flowchart of the road obstruction detection method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 101: Obtain the pre-disaster and post-disaster images to be detected.

[0030] Pre-disaster imagery refers to remote sensing images acquired before a disaster occurs, reflecting the distribution of road networks and surrounding features in a target area under normal conditions.

[0031] Post-disaster imagery refers to remote sensing images acquired after a disaster, reflecting the surface condition and road damage in the target area.

[0032] Two images, one before and one after the disaster, are acquired from remote sensing data sources (such as high-resolution optical satellite imagery or UAV aerial imagery). The pre-disaster image is typically a high-quality image with no or low cloud cover, selected from the most recent period before the disaster. The post-disaster image is usually acquired within hours of the disaster through emergency observation missions. The two images undergo geometric registration to ensure accurate spatial alignment of the same ground features in different temporal images. Furthermore, the original images are preprocessed, including radiometric calibration, atmospheric correction, cropping, and normalization, to eliminate the impact of differences in imaging conditions on subsequent feature extraction. The acquired pre-disaster and post-disaster images have the same spatial resolution, spatial extent, and coordinate system, forming a dual-temporal image pair, which is then input into a target detection model for the identification and detection of road blockage areas.

[0033] Step 102: Input the pre-disaster image and the post-disaster image into the target detection model to obtain the change detection results of the road blockage area output by the target detection model.

[0034] The change detection result of road blockage area refers to the output data of the spatial location and range of road network blockage caused by natural disasters (such as landslides, collapses, debris flows, floods, etc.) in the input image, which is usually presented in the form of a binary change mask.

[0035] The object detection model is a pre-trained deep learning network model used to receive pre-disaster and post-disaster images and output the detection results of changes in road blockage areas. (Reference) Figure 2The model employs an encoder-decoder architecture, comprising a dual-branch feature extraction backbone network and a cross-guided neck network. The dual-branch feature extraction backbone network includes a CNN (Convolutional Neural Network) branch and a ViT (Vision Transformer) branch. The CNN branch uses two independent ResNet-18 (Residual Network with 18 layers) branches to extract local detail features from pre-disaster and post-disaster images, respectively. The ViT branch uses InternViT-Large with shared weights to extract shared global semantic features from the two temporal images. The cross-guided neck network consists of a cascaded temporal-specific enhancement module, a bidirectional cross-guided fusion module, and a spatially gated enhancement module, used for differential enhancement, collaborative fusion, and gated enhancement of the two temporal features. The model also includes a decoder, typically employing a UPerHead (Unified Perceptual Parsing Head), which performs multi-scale context aggregation on the enhanced fused features and outputs a change mask for the road blockage area. During the training phase, the model undergoes supervised learning using a large number of labeled bi-temporal image pairs and corresponding real change masks to optimize network parameters. During the inference phase, the model parameters are fixed, and only forward computation is performed, mapping the input pre-disaster and post-disaster images to change detection results. This target detection model operates end-to-end, requiring no manual feature design or intermediate steps, achieving direct mapping from bi-temporal images to road blockage areas.

[0036] The target detection model includes a cross-guided neck network, which comprises: ①The Temporal Specific Module (TSM) is used to apply edge structure enhancement to the pre-disaster local features of pre-disaster images, apply spatial region enhancement to the post-disaster local features of post-disaster images, and extract temporal difference priors based on the differences between pre-disaster and post-disaster local features.

[0037] It should be understood that pre-disaster local features refer to feature maps representing local details of roads and surrounding features, including fine-grained spatial information such as road edges, textures, and shapes. Post-disaster local features refer to feature maps representing local details of blocked areas such as landslides and collapses, including fine-grained spatial information such as the edges, textures, and shapes of the blocked areas. Edge structure enhancement refers to applying attention operations based on edge detection operators to pre-disaster local features to strengthen the linear structure and boundary response of roads and suppress noise responses in non-road areas. Spatial region enhancement refers to applying attention operations based on global spatial context modeling to post-disaster local features to strengthen the overall consistency response of blocked areas and suppress the response of discrete noise points. Temporal difference prior refers to feature maps extracted based on pixel-by-pixel differences between pre-disaster and post-disaster local features, used to represent the intensity of change between two temporal images.

[0038] In the temporal-specific enhancement module, edge attention operations are used to enhance the linear structural response of roads in response to their elongated linear characteristics before disasters. Spatial attention operations are used to enhance the regional consistency of blocked areas in response to the clumpy characteristics of landslides and collapses after disasters. At the same time, prior temporal differences reflecting the changes in location and intensity are extracted by subtracting pixel by pixel from local features of the two temporal phases and performing convolution transformation.

[0039] ② The Simplified Cross Guided Fusion (SCGF) module is used to perform mutually constrained collaborative fusion of shared global semantic features of pre-disaster and post-disaster images with temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of enhanced pre-disaster local features, enhanced post-disaster local features and temporal difference priors.

[0040] It should be understood that shared global semantic features refer to multi-scale semantic feature maps with global context modeling capabilities extracted from pre-disaster and post-disaster images through a shared weighted visual Transformer backbone, used to capture the topological structure of the road network and the overall spatial distribution information of the landslide area.

[0041] In the bidirectional cross-guided fusion module, spatial guidance weights for temporal-specific fusion features are generated using shared global semantic features to achieve semantic filtering. At the same time, spatial guidance weights for shared global semantic features are generated using temporal-specific fusion features to achieve detail compensation. The results of the two bidirectional guidance are fused to output a cross-guided fusion feature that simultaneously possesses local details and global semantic information.

[0042] ③ The Supervised Gate Module (SGM) is used to enhance the cross-guided fusion features based on the enhanced pre-disaster local features, enhanced post-disaster local features, and temporal difference priors generated by the spatial gating map, and output the enhanced fusion features; the change detection results of the road blockage area are generated based on the enhanced fusion features.

[0043] In the spatial gating enhancement module, the enhanced pre-disaster local features, enhanced post-disaster local features, and temporal difference priors are spliced ​​together. A gating weight map reflecting the spatial distribution of the changed area is generated by the gating predictor. The gating weight map is then used to perform non-destructive additive enhancement on the cross-guided fusion features, and the enhanced fusion features with enhanced response in the changed area and unchanged background response are output.

[0044] The enhanced fused features are input into the decoder, which, through multi-scale context aggregation and pixel-level classification, outputs a binary transformation mask of the same size as the input image. Each pixel value in the mask identifies whether the location belongs to a road blockage area.

[0045] In one embodiment, pre-disaster and post-disaster images, after preprocessing such as geometric registration, radiometric correction, and normalization, are subjected to multi-scale feature extraction through a dual-branch feature extraction backbone network of a target detection model. The target detection model performs temporal-specific enhancement, bidirectional cross-guided fusion, and spatial gating enhancement on the dual-temporal features in a cross-guided neck network, and finally outputs a change detection result of the same size as the input image through a decoder. This change detection result is typically presented in the form of a binary change mask, where changed pixels (such as areas where roads are blocked) are marked as 1 (white), and unchanged pixels (such as normal roads and surrounding background) are marked as 0 (black). The change mask identifies whether each pixel location belongs to a road blockage area, thereby obtaining information such as the spatial location, geometry, and coverage of the blockage area. This change detection result can be directly used for subsequent applications such as generating thematic maps of road blockages, calculating the length and area of ​​blocked roads, and planning emergency rescue routes.

[0046] The road blockage detection method provided in this invention employs a temporal-specific enhancement module to apply edge structure enhancement to pre-disaster images to highlight the linear features of slender roads, and a spatial region enhancement module to apply spatial region enhancement to post-disaster images to strengthen the blocky consistency of landslides and collapses. It also extracts prior information on dual-temporal differences, effectively solving the feature mismatch problem caused by significant differences in target morphology between pre- and post-disaster images. A bidirectional cross-guided fusion module achieves mutual constraint and collaborative fusion of local detail features and global semantic features, avoiding feature fragmentation caused by simple fusion. A spatial gating enhancement module uses temporal difference priors to generate gating weights for non-destructive additive enhancement of the fused features, suppressing background noise while increasing the response intensity of low-proportion change areas. The synergistic effect of these three components improves the accuracy of road blockage detection.

[0047] Based on the above embodiments, the target detection model further includes a dual-branch feature extraction backbone network, which includes: The first feature extraction branch is used to extract the pre-disaster local features of the pre-disaster image and the post-disaster local features of the post-disaster image. The second feature extraction branch is used to extract the shared global semantic features of the pre-disaster images and the post-disaster images.

[0048] Before the cross-guided neck network, the target detection model sets up two parallel feature extraction branches. The first feature extraction branch uses two independent convolutional neural networks to extract multi-scale local detail features from pre-disaster and post-disaster images, respectively. The second feature extraction branch uses a visual Transformer with shared weights to extract global semantic features from dual-temporal images. Together, they constitute the encoder part, providing complementary multi-level feature representations for the subsequent neck network.

[0049] refer to Figure 3 The first feature extraction branch employs two independent ResNet-18 convolutional neural networks to process pre-disaster and post-disaster images, respectively. The parameters of the two networks are independent of each other to avoid temporal confusion and ensure that each can be optimized for the characteristics of its own time phase.

[0050] Each ResNet-18 network contains four residual stages. The input image (256×256×3) passes through the initial convolutional layer, the max pooling layer, and the four residual stages in sequence, outputting feature maps at four scales: Stage 0 is (spatial resolution 1 / 4); Stage 1 is (Spatial resolution 1 / 8); Stage 2 is (Spatial resolution 1 / 16); Stage 3 is (Spatial resolution 1 / 32); These multi-scale feature maps gradually expand the receptive field and increase the number of channels from low to high. The shallow features (high resolution) retain detailed information such as road edges and textures, while the deep features (low resolution) contain richer semantic information. To align with the feature dimensions of the ViT branches, the number of channels in each stage is uniformly projected to 1024 using 1×1 convolutions.

[0051] The second feature extraction branch uses the InternViT-Large network with shared weights as the backbone. It uses the same network parameters for feature extraction on both pre-disaster and post-disaster images to ensure that the features of both time phases are in a consistent semantic space. InternViT-Large models long-distance dependencies in images through a self-attention mechanism, which can effectively capture the overall topological connectivity of the road network and the large-scale spatial distribution of landslides, collapses and other blocking areas.

[0052] In terms of training strategy, the ViT backbone uses pre-trained parameters and keeps them frozen (i.e., it does not participate in gradient updates), training only task-related modules such as the neck network and decoder. This design can stably utilize global context modeling capabilities under limited sample conditions, while controlling training costs and the risk of overfitting.

[0053] In this embodiment of the invention, local details before and after a disaster are extracted independently through CNN branches, while maintaining fine-grained information such as road edges; global semantics in two phases are extracted through ViT branches with shared weights, ensuring consistency in the feature space; the two work in parallel to provide multi-level features with complementary information for the subsequent neck network, thereby improving the joint perception capability of slender roads and blocky landslides.

[0054] Based on the above embodiments, the time-phase specific enhancement module is specifically used for: The enhanced pre-disaster local features are obtained by performing an edge attention operation on the pre-disaster local features. The enhanced post-disaster local features are obtained by performing a spatial attention operation on the post-disaster local features. The difference between the pre-disaster local features and the post-disaster local features is calculated pixel by pixel to obtain the difference features. The difference features are then transformed through convolution to obtain the temporal difference prior.

[0055] The temporal-specific enhancement module is the first-level module of the cross-guided neck network, used to address the feature mismatch caused by differences in the geometry of pre-disaster roads and post-disaster blocked areas. (Reference) Figures 4-5 This module contains three parallel processing paths: (1) Pre-enhance path (edge ​​structure enhancement): Targeting pre-disaster local features, multiple stacked first convolutions (such as 3×3 convolutions) and an edge attention module based on the Sobel operator are used to perform edge detection and feature enhancement on slender roads in pre-disaster images. For example, the Sobel operator is used to extract the gradient responses in the horizontal and vertical directions to generate an edge feature map; an edge attention weight map is generated through 1×1 convolutions and Sigmoid activation. The weight map is then multiplied element-wise with the original pre-disaster local features to obtain the enhanced pre-disaster local features. This operation enhances the linear structural response of the roads while suppressing the response of non-road areas.

[0056] (2) Post-enhance path (spatial region enhancement): For post-disaster local features, multiple stacked second convolutions (such as 5×5 convolutions) and a spatial attention module based on global average pooling are used to enhance the overall consistency of landslide, collapse, and other blocking areas in post-disaster images. For example, global average pooling is performed on the post-disaster local features to obtain a global context feature vector; a spatial attention weight map is generated through 1×1 convolution and sigmoid activation, and the weight map is multiplied element-wise with the original post-disaster local features to obtain the enhanced post-disaster local features. This operation strengthens the internal consistency of the blocky blocking areas and suppresses isolated noise points.

[0057] (3) diff_extractor path (temporal difference extraction): Extracts temporal difference priors based on pixel-by-pixel differences between the original pre-disaster local features and post-disaster local features. For example, the difference between the two temporal local features (usually the difference between post-disaster and pre-disaster mitigation features or absolute value differences) is obtained to obtain a difference feature map; the difference feature map is transformed by convolution operations (such as stacked 3×3 convolution, batch normalization, and ReLU activation) to obtain a temporal difference prior feature map. This difference prior reflects the location and intensity information of changes between the two temporal phases, providing prior knowledge for the spatial gating enhancement of the subsequent SGM module to locate the change area.

[0058] The enhanced pre-disaster local features, enhanced post-disaster local features, and temporal difference priors are concatenated and then compressed using 1×1 convolution to output temporal-specific fusion features, which serve as the input to the bidirectional cross-guided fusion module.

[0059] In one embodiment, in RoadGuard-Net (Road Guard Network), Neck does not perform a one-time stitching of bi-temporal features, but rather performs continuous transformations at each scale, including temporal-specific extraction, cross-branch fusion, and explicit gating enhancement. Let the... Each scale input is , and (All channels are 1024). TSM is responsible for first separating the pre-disaster and post-disaster features and then explicitly injecting the differences into the subsequent fusion process.

[0060] TSM is composed of three paths: pre_enhance, post_enhance, and diff_extractor. pre_enhance targets the fine-grained road structure before the disaster and employs... Convolutional stacking combined with the EdgeAttentionModule enhances edge continuity, and Sobel-based edge attention is introduced. Specifically, for input features... First, perform grouped convolutional Sobel edge detection to obtain dual-channel stacked edge features. It is compressed into an attention map through a fusion layer. Finally, element-wise weighted average is performed. The goal of this process is to enhance the linear structural response of slender roads in pre-disaster imagery. `post_enhance` targets post-disaster landslide and collapse areas, employing 5×5 convolution stacking combined with a SpatialAttentionModule to enhance regional consistency; `diff_extractor` extracts change-discriminating priors from the subtracted bi-temporal features. The three outputs are denoted as follows: ; in, Indicates scale index. Indicates the first Enhanced local features before the disaster at various scales Indicates the first Enhanced local post-disaster features at various scales Indicates the first Priors for temporal differences at various scales, Indicates the first Local pre-disaster features at various scales Indicates the first Local post-disaster features at various scales This represents the pre-disaster enhancement function. This represents the post-disaster enhancement function. This represents the difference extraction function. This indicates a difference operation.

[0061] The Sobel operator performs first-order difference on the local neighborhood using a fixed gradient kernel, which can highlight locations of abrupt gray-level changes while preserving edge orientation information. For pre-disaster road scenarios, this type of gradient response exhibits high consistency with slender linear structures, making it suitable as a priori signal for edge attention.

[0062] The three types of complementary information are compressed into a unified representation using a temporal-specific enhancement module: ; in, Indicates the first Temporal specific fusion features at various scales, This represents a 1×1 convolution (compression). This indicates channel splicing (fusion). The purpose of this step is not simply dimensionality reduction, but to map pre-disaster structural clues, post-disaster regional clues, and temporal difference clues to the same discriminative subspace, providing an alignable CNN semantic basis for subsequent cross-guidance.

[0063] It should be noted that TSM only applies to the CNN branch and not directly to the ViT branch, mainly based on two considerations. First, the Sobel edge enhancement and local spatial attention in TSM are local structural priors, suitable for application to convolutional features that preserve pixel neighborhood topology; if applied directly to ViT token features, it easily weakens its global relation modeling advantage. Second, in the setup of this invention, the ViT backbone uses shared weights and remains frozen, aiming to provide stable and consistent global semantic coordinates for both temporal phases; if additional temporal-specific perturbations are applied to ViT in the Neck, it will increase the risk of semantic space drift and weaken the alignment stability with the bidirectional guidance of SCGF. Based on the above division of labor, TSM is responsible for enhancing the detailed difference representation on the CNN side, while the ViT side maintains global semantic consistency, and SCGF completes the complementary fusion of the two types of information.

[0064] This invention addresses the feature mismatch problem caused by morphological differences by strengthening the linear structure of roads before disasters through edge attention and strengthening the blocky consistency of blocked areas after disasters through spatial attention. At the same time, it provides prior location information for low-proportion change areas by extracting temporal difference priors through pixel-by-pixel subtraction, thereby improving detection accuracy.

[0065] Based on the above embodiments, the bidirectional cross-guided fusion module is specifically used for: The shared global semantic features are used to generate a first spatial guiding weight for the temporal-specific fusion features, and the first spatial guiding weight is used to weight the temporal-specific fusion features. The temporal-specific fusion features are used to generate a second spatial guiding weight for the shared global semantic features, and the shared global semantic features are weighted using the second spatial guiding weight; The weighted temporal-specific fusion features are fused with the weighted shared global semantic features to output the cross-guided fusion features.

[0066] In the bidirectional cross-guided fusion module, spatial attention weights are generated from two directions: shared global semantic features and temporal-specific fusion features, respectively, to achieve bidirectional interactive weighting of features. Then, the two weighted features are fused to obtain cross-guided fusion features that take into account both local details and global semantics.

[0067] refer to Figure 5 SCGF achieves mutual constraint and collaborative fusion of CNN local features and ViT global semantic features through the following steps: (1) Utilizing shared global semantic features Generate phase-specific fusion features First Space Guiding Weight Specifically, A spatial weight map is generated using 1×1 convolution and the Sigmoid activation function. ,in, This represents a 1×1 convolution operation. This represents the Sigmoid activation function. Then, this weight map is used to perform element-wise weighting of the temporal-specific fusion features: This step leverages ViT's global semantic understanding capabilities to generate spatial attention, suppressing local noise responses in CNN features and highlighting regions relevant to semantic changes.

[0068] (2) Utilizing temporal-specific fusion features Generate shared global semantic features Second Space Guiding Weight Specifically, A spatial weight map is generated using 1×1 convolution and the Sigmoid activation function. Then, the shared global semantic features are weighted element-wise using this weight graph: This step leverages the local detail perception capabilities of CNNs to generate spatial attention, improving the issue of blurred boundaries in ViT features and enhancing the response to fine-grained information such as road edges.

[0069] (3) The weighted temporal-specific fusion features Weighted shared global semantic features The data is concatenated along the channel dimension and then adaptively recombined using a 1×1 convolution to output cross-guided fusion features: ; in, This indicates cross-guided fusion characteristics.

[0070] Through the aforementioned bidirectional guidance mechanism, the SCGF module achieves deep interaction between local details and global semantics. The output cross-guided fusion features retain the local structural resolution of CNN and inherit the global semantic organization capability of ViT, effectively avoiding the feature fragmentation problem caused by simple fusion, and providing high-quality feature input for the spatial gating enhancement of the subsequent SGM module.

[0071] This invention achieves semantic filtering of CNN by ViT and detail compensation of ViT by CNN through bidirectional generation of spatial guided weights, so that the fused features have both local details and global semantics, avoiding feature fragmentation and improving detection accuracy.

[0072] Based on the above embodiments, the space gating enhancement module is specifically used for: The enhanced pre-disaster local features, the enhanced post-disaster local features, and the temporal difference prior are spliced ​​together to obtain spliced ​​features; The stitching features are processed by a gating predictor and activated by an activation function to generate a spatial gating map; Using the spatial gating graph, the cross-guided fusion feature is non-destructively additively enhanced to obtain the enhanced fusion feature.

[0073] It should be understood that a spatial gating map refers to a single-channel weight map of the same size as the input features, generated in the spatial gating enhancement module based on enhanced pre-disaster local features, enhanced post-disaster local features, and temporal difference priors, through a gating predictor and a sigmoid activation function. Each pixel value is between 0 and 1, used to characterize the gating weight of the location as belonging to a road blockage area. The larger the value, the more likely the location is to be a change area.

[0074] In the spatial gating enhancement module, the three features are concatenated along the channel dimension. A gating weight map reflecting the spatial distribution of the changing region is generated by a gating predictor and a sigmoid activation function. This gating weight map is then used to additively enhance the cross-guided fusion features, resulting in an enhanced fusion feature where the response of the changing region is strengthened while the background response remains unchanged.

[0075] refer to Figure 5 SGM is the third-level module of the cross-guided neck network, used to address class imbalance caused by the low pixel proportion in variable regions. This module achieves selective enhancement of variable regions through the following steps: (1) The enhanced pre-disaster local features, the enhanced post-disaster local features, and the temporal difference prior are spliced ​​together in the channel dimension to obtain the spliced ​​features. : ; The three features have the same number of channels (1024 each), and the number of channels in the stitched feature is 3072. This stitched feature integrates the linear structure information of the road before the disaster, the clumping information of the blocked area after the disaster, and the location and intensity information of the dual temporal changes, providing rich discrimination criteria for gating map generation.

[0076] (2) Input the spliced ​​features into the gating predictor Then, after activation by the Sigmoid activation function, a spatial gating graph is generated. : ; The gated predictor typically consists of convolutional layers, batch normalization layers, ReLU activation functions, and 1×1 convolutional layers, used to learn the spatial distribution patterns of changing regions from the stitched features. The Sigmoid activation function compresses the output values ​​to between 0 and 1, resulting in the spatially gated map. The gating map is the same size as the input feature (e.g., H×W). Each pixel value represents the gating weight of the location as belonging to a changing region. The closer the value is to 1, the more likely the location is to be a changing region.

[0077] (3) Using the generated spatial gating graph Cross-guided fusion features Perform additive enhancement: ; in, Indicates the first Enhanced fusion features of the neck network output guided by layer cross-linking. When When it is close to 0 (non-changing region). The background representation remains unchanged; when When it approaches 1 (the area of ​​change). The response in the changed region was boosted by about one time.

[0078] The above design also introduces additional supervision signals during the training phase, such as spatial gating graphs. Supervised by the binary cross-entropy loss of the true change mask, the gated map can more accurately focus on the changing regions. Specifically, during model training, the binary cross-entropy loss between the generated spatial gated map at each scale and the corresponding true change mask is calculated and used as an auxiliary supervision signal to jointly optimize the network parameters with the main segmentation loss. During inference, no additional gated prediction branch is introduced; the trained gated predictor is used directly to generate the spatial gated map, and additive enhancement is applied to the fused features.

[0079] During the training phase, spatial gating diagrams Masked by real changes Explicit supervision. Specifically, for each scale (There are 4 scales in total, corresponding to the feature map sizes of the 4 stages output by ResNet-18), masking the real changes. Downsampling to With the same spatial resolution, we obtain Then calculate. and Binary Cross Entropy (BCE) loss between: ; Through this explicit supervision, the spatial gating graph can learn the accurate spatial distribution of changing regions, providing reliable gating weights for subsequent additive enhancement.

[0080] Gated loss With the main splitting loss (cross-entropy loss) Together they constitute a joint optimization objective: ; in, Represents the weighting coefficients that balance the two loss terms; Cross-entropy loss is defined as follows: ; in, Indicates the image height. Indicates the image width. Indicates the total number of pixels. Indicates the category (change / non-change); This represents the true label (one-hot encoding), i.e., the position. This belongs to the category The actual label, with a value of 0 or 1; This represents the model's predicted probability, i.e., the model's predicted location. This belongs to the category The probability of.

[0081] Through the above steps, the SGM module improves the response intensity of low-proportion change regions without sacrificing background stability, effectively mitigating the impact of class imbalance on detection accuracy.

[0082] This invention generates a spatial gating map by using a priori guidance based on temporal differences, and performs non-destructive additive enhancement on the fused features. While maintaining background stability, it directionally enhances low-proportion change regions, effectively alleviating the class imbalance problem and improving detection recall.

[0083] Based on the above embodiments, the target detection model further includes a decoder, which is used to perform multi-scale context aggregation on the enhanced fusion features to generate the change detection result of the road blocking area.

[0084] It should be understood that multi-scale context aggregation refers to the process in the decoder where the deepest features are pooled with different receptive fields through a pyramid pooling module to capture multi-scale contextual information from local to global. Then, the deep semantic information is passed from top to bottom to the shallow features through a feature pyramid network, achieving cross-scale feature fusion. This enables the final features to have both wide-range semantic perception capabilities and fine spatial localization capabilities.

[0085] In the target detection model, a decoder module is set up to receive the multi-scale enhanced and fused features output by the cross-guided neck network. The context information of different scales is aggregated through pyramid pooling, and the features of each layer are fused from top to bottom through the feature pyramid network. Finally, a binary transformation mask that identifies the spatial location and extent of the road blockage area is output.

[0086] The embodiments of the present invention improve the ability to identify blocking regions at different scales and the accuracy of spatial positioning by multi-scale context aggregation and fusion of deep semantics and shallow details.

[0087] Based on the above embodiments, the decoder is specifically used for: The enhanced fusion features are multi-scale pooled by the pyramid pooling module to obtain multi-scale pooled features. The multi-scale pooling features and the enhanced fusion features of each layer are fused from top to bottom using a feature pyramid network to obtain multi-scale fusion features. Based on the multi-scale fusion features, the change detection results of the road blockage area are generated.

[0088] The Pyramid Pooling Module (PPM) is a feature processing module in the decoder that simultaneously applies multiple pooling operations of different scales (such as 4×4, 5×5, 7×7, etc.) to the deepest enhanced and fused features, upsamples each pooling result to the original resolution, and then concatenates them to capture multi-scale contextual information from local to global.

[0089] refer to Figure 6 PPM performs the following operations on the deepest enhanced post-fusion features (with the smallest spatial resolution and the highest degree of semantic abstraction): (1) Apply multiple pooling operations of different scales simultaneously. Taking pooling scales of 4, 5, and 7 as an example: Divide the feature map into a 4×4 grid and perform average pooling (or max pooling) on ​​each grid to obtain a 4×4 pooled feature; divide the feature map into a 5×5 grid and perform average pooling (or max pooling) on ​​each grid to obtain a 5×5 pooled feature; divide the feature map into a 7×7 grid and perform average pooling (or max pooling) on ​​each grid to obtain a 7×7 pooled feature.

[0090] (2) Upsample the pooled features at each scale to the same spatial resolution as the original deepest features through bilinear interpolation (or transposed convolution), and then concatenate them along the channel dimension to form a feature representation containing multi-scale contextual information: ; in, Represents the original deepest features. , , , This represents pooling operations with scales of 1, 2, 3, and 6.

[0091] By incorporating contextual information at different scales into the features processed by PPM, each location is integrated: small-scale pooling allows the model to focus on local details, such as the precise boundaries of small landslides; large-scale pooling allows the model to perceive the global scene, such as the overall extent of large landslides. This design enables the model to perceive blocking areas of different sizes simultaneously, enhancing scale robustness.

[0092] refer to Figure 6 The Feature Pyramid Network (FPN) is a feature fusion module in the decoder that upsamples deep features layer by layer from top to bottom and adds them to the shallow features of the corresponding layer, so as to realize the transmission of high-level semantic information to the shallow layer and enable the final features to have both semantic discrimination ability and spatial positioning accuracy.

[0093] FPN performs top-down fusion of the multi-scale pooling features output by PPM with the enhanced post-fusion features output by each layer of the encoder (or neck network): (1) Starting from the deepest layer (semantically strongest and spatially coarsest), process layer by layer upwards: upsample the deep features to 2 times the resolution through bilinear interpolation (or transposed convolution); add and fuse the upsampled deep features with the shallow features of the corresponding layer (from the encoder or neck network).

[0094] (2) Before each layer is fused, the shallow features are usually adjusted by 1×1 convolution to make them consistent with the number of channels of the upsampled deep features, so as to facilitate element-by-element addition: ; in, Indicates the first The fusion features of the FPN output. Indicates an upsampling operation. Indicates the first The fusion features of the FPN output. Through top-down fusion using FPN: semantic information from deep features is passed to shallow features, and spatial details from shallow features (such as the precise boundaries of landslides) are added to deep features. The final multi-scale fused features have both semantic discrimination capabilities and precise spatial positioning accuracy.

[0095] Finally, FPN features at different scales are upsampled and concatenated, and a change prediction mask is generated through 1×1 convolution: ; in, Indicates a change mask. These represent the output features of the 0th, 1st, 2nd, and 3rd layers of the FPN, respectively.

[0096] The embodiments of the present invention improve the recognition accuracy and boundary localization capability of blocking regions at different scales by using multi-scale pooling and top-down fusion to make features have both global semantics and local details.

[0097] To further explain the road obstruction detection method proposed in this invention, please refer to the following embodiments.

[0098] Model evaluation was performed using the BlockedRoad and LEVIR-CD datasets. To comprehensively evaluate the performance of RoadGuard-Net, seven representative change detection methods were selected as baselines, covering CNN, Transformer, and hybrid architectures.

[0099] CNN-based methods: STANet proposed a spatiotemporal attention network, which enhances the ability to focus on changing regions through spatial and channel attention mechanisms, achieving performance improvements on the LEVIR-CD dataset. BAN designed a dual attention network, combining channel and spatial attention to improve feature representation capabilities, and introduced an adversarial training mechanism to enhance model robustness.

[0100] Transformer-based methods: BIT first introduced Transformer into change detection, representing bi-temporal images as semantic tokens and modeling spatiotemporal context in a compact token space, achieving performance improvements with lower computational cost. ChangeFormer proposed a Transformer Siamese architecture, unifying the hierarchical Transformer encoder and MLP decoder into a single framework, enhancing the ability to model multi-scale long-distance dependencies. MetaChanger emphasized the importance of feature interaction, proposing a general architecture composed of alternating interaction layers to achieve zero-parameter interaction through feature exchange.

[0101] Hybrid architecture approaches: VITP-L combines the advantages of Vision Transformer and CNN, extracting global semantics and local detail features simultaneously through a dual-branch architecture. DCSI-UNet enhances feature fusion through a deep channel spatial interaction module and employs a dual-stream UNet architecture to achieve multi-scale feature aggregation. TinyCD designs a lightweight change detection network that significantly reduces computational costs while maintaining high accuracy, making it suitable for resource-constrained scenarios.

[0102] Table 1 shows the performance comparison between RoadGuard-Net and existing methods on the BlockedRoad dataset. The comparison baselines include CNN-based methods (STANet, BAN), Transformer-based methods (ChangeFormer, MetaChanger), hybrid architecture methods (VITP-L, DCSI-UNet), and lightweight methods (TinyCD).

[0103] Table 1 Performance comparison on the BlockedRoad dataset RoadGuard-Net achieved state-of-the-art performance on the BlockedRoad dataset. Compared to the baseline VITP-L, IoU improved from 60.51% to 66.6% (+6.09%), and mIoU improved from 79.84% to 83.01% (+3.17%). This result is consistent with the conclusions of recent studies on the collaborative modeling of local details and global semantics: relying on a single branch usually makes it difficult to simultaneously take into account boundary continuity and region consistency [12-13,16-17]. Figure 7 Visual comparisons are provided for typical scenarios.

[0104] Figure 7 Visual comparison of the BlockedRoad dataset. From left to right: pre-disaster image, post-disaster image, labels, and prediction results from each baseline and RoadGuard-Net. RoadGuard-Net shows greater stability in terms of the continuity of thin-line roads and the integrity of blocked areas.

[0105] Table 2 shows the performance comparison of RoadGuard-Net on the LEVIR-CD dataset, which is used to evaluate the transferability of the method to general change detection tasks.

[0106] Table 2 Performance comparison on the LEVIR-CD dataset RoadGuard-Net also achieved the best or tied best on LEVIR-CD (IoU=83.6%, mIoU=91.41%), indicating that the method is not only applicable to road blockage scenarios. Compared with data distributions dominated by building changes, LEVIR-CD has more regular target shapes and more stable semantic contrasts, so the differences between the methods converged overall, a phenomenon consistent with common observations in public benchmarks [9-11,15-16]. Figure 8 The results show that RoadGuard-Net can effectively suppress boundary overflow and discrete false positives while maintaining high recall.

[0107] Figure 8 Visual comparison of the LEVIR-CD dataset. RoadGuard-Net maintains good predictive consistency between target boundaries and region integrity.

[0108] The differences between methods on LEVIR-CD are relatively narrowing, which is related to the fact that their target shapes are more regular and their semantic contrasts are clearer. In contrast, the thin line structure and irregular disaster areas coexist in BlockedRoad, thus better reflecting the model's ability to adapt to heterogeneous change patterns.

[0109] The road obstruction detection device provided by the present invention is described below. The road obstruction detection device described below can be referred to in correspondence with the road obstruction detection method described above.

[0110] refer to Figure 9 The road obstruction detection device provided by the present invention includes: The acquisition module 901 is used to acquire the pre-disaster and post-disaster images to be detected; The road blockage detection module 902 is used to input the pre-disaster image and the post-disaster image into a target detection model, and obtain the change detection results of the road blockage area output by the target detection model; wherein, the target detection model includes an inter-guide neck network, and the inter-guide neck network includes: The temporal-specific enhancement module is used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference priors based on the differences between the pre-disaster local features and the post-disaster local features. The bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster images and the post-disaster images with the temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference prior. The spatial gating enhancement module is used to enhance the cross-guidance fusion features based on the enhanced pre-disaster local features, the enhanced post-disaster local features, and the spatial gating map generated by the temporal difference prior, and output the enhanced fusion features; the change detection result of the road blockage area is generated based on the enhanced fusion features.

[0111] The road blockage detection device provided in this invention employs a temporal-specific enhancement module to apply edge structure enhancement to pre-disaster images to highlight the linear features of slender roads and a spatial region enhancement module to apply spatial region enhancement to post-disaster images to strengthen the blocky consistency of landslides and collapses. It also extracts prior information on dual-temporal differences, effectively solving the feature mismatch problem caused by significant differences in target morphology between pre- and post-disaster images. A bidirectional cross-guided fusion module achieves mutual constraint and collaborative fusion of local detail features and global semantic features, avoiding feature fragmentation caused by simple fusion. A spatial gating enhancement module uses temporal difference priors to generate gating weights for non-destructive additive enhancement of the fused features, suppressing background noise while increasing the response intensity of low-proportion change areas. The synergistic effect of these three components improves the accuracy of road blockage detection.

[0112] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10As shown, the electronic device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute a road blockage detection method, which includes: acquiring pre-disaster and post-disaster images to be detected; inputting the pre-disaster and post-disaster images into a target detection model, and obtaining the road blockage area change detection result output by the target detection model; wherein the target detection model includes a cross-guided neck network, and the cross-guided neck network includes: a temporal-specific enhancement module, used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference based on the difference between the pre-disaster and post-disaster local features. The system includes a two-way cross-guided fusion module, which performs mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster and post-disaster images with temporal-specific fusion features, and outputs cross-guided fusion features. The temporal-specific fusion features are fusion features of enhanced pre-disaster local features, enhanced post-disaster local features, and the temporal difference prior. A spatial gating enhancement module is used to enhance the cross-guided fusion features based on the spatial gating map generated by the enhanced pre-disaster local features, the enhanced post-disaster local features, and the temporal difference prior, and outputs enhanced fusion features. The change detection result of the road blockage area is generated based on the enhanced fusion features.

[0113] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] On the other hand, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the road blockage detection method provided by the above methods. The method includes: acquiring pre-disaster images and post-disaster images to be detected; inputting the pre-disaster images and the post-disaster images into a target detection model, and acquiring the change detection result of the road blockage area output by the target detection model; wherein the target detection model includes an inter-guided neck network, the inter-guided neck network including: a temporal-specific enhancement module, used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster images, apply spatial region enhancement to the post-disaster local features of the post-disaster images, and based on the pre-disaster local features and the post-disaster images... The system extracts temporal difference priors from the differences between local features; a bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster and post-disaster images with temporal-specific fusion features, and outputs cross-guided fusion features; the temporal-specific fusion features are fusion features of enhanced pre-disaster local features, enhanced post-disaster local features and the temporal difference priors; a spatial gating enhancement module is used to enhance the cross-guided fusion features based on the spatial gating map generated by the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference priors, and output the enhanced fusion features; the change detection results of the road blockage area are generated based on the enhanced fusion features.

[0115] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0116] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting road blockage, characterized in that, include: Acquire pre-disaster and post-disaster images to be inspected; The pre-disaster and post-disaster images are input into a target detection model to obtain the change detection results of road blockage areas output by the target detection model; wherein, the target detection model includes an inter-guide neck network, and the inter-guide neck network includes: The temporal-specific enhancement module is used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference priors based on the differences between the pre-disaster local features and the post-disaster local features. The bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster images and the post-disaster images with the temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference prior. The spatial gating enhancement module is used to enhance the cross-guidance fusion features based on the enhanced pre-disaster local features, the enhanced post-disaster local features, and the spatial gating map generated by the temporal difference prior, and output the enhanced fusion features; the change detection result of the road blockage area is generated based on the enhanced fusion features.

2. The road obstruction detection method according to claim 1, characterized in that, The time-phase specific enhancement module is specifically used for: The enhanced pre-disaster local features are obtained by performing an edge attention operation on the pre-disaster local features. The enhanced post-disaster local features are obtained by performing a spatial attention operation on the post-disaster local features. The difference between the pre-disaster local features and the post-disaster local features is calculated pixel by pixel to obtain the difference features. The difference features are then transformed through convolution to obtain the temporal difference prior.

3. The road obstruction detection method according to claim 1, characterized in that, The bidirectional cross-guided fusion module is specifically used for: The shared global semantic features are used to generate a first spatial guiding weight for the temporal-specific fusion features, and the first spatial guiding weight is used to weight the temporal-specific fusion features. The temporal-specific fusion features are used to generate a second spatial guiding weight for the shared global semantic features, and the shared global semantic features are weighted using the second spatial guiding weight; The weighted temporal-specific fusion features are fused with the weighted shared global semantic features to output the cross-guided fusion features.

4. The road obstruction detection method according to claim 1, characterized in that, The space gating enhancement module is specifically used for: The enhanced pre-disaster local features, the enhanced post-disaster local features, and the temporal difference prior are spliced ​​together to obtain spliced ​​features; The stitching features are processed by a gating predictor and activated by an activation function to generate a spatial gating map; Using the spatial gating graph, the cross-guided fusion feature is non-destructively additively enhanced to obtain the enhanced fusion feature.

5. The road obstruction detection method according to claim 1, characterized in that, The target detection model further includes a dual-branch feature extraction backbone network, which comprises: The first feature extraction branch is used to extract the pre-disaster local features of the pre-disaster image and the post-disaster local features of the post-disaster image. The second feature extraction branch is used to extract the shared global semantic features of the pre-disaster images and the post-disaster images.

6. The road blockage detection method according to claim 1, characterized in that, The target detection model also includes a decoder, which is used to perform multi-scale context aggregation on the enhanced fusion features to generate change detection results for the road blockage area.

7. The road obstruction detection method according to claim 6, characterized in that, The decoder is specifically used for: The enhanced fusion features are multi-scale pooled by the pyramid pooling module to obtain multi-scale pooled features. The multi-scale pooling features and the enhanced fusion features of each layer are fused from top to bottom using a feature pyramid network to obtain multi-scale fusion features. Based on the multi-scale fusion features, the change detection results of the road blockage area are generated.

8. A road obstruction detection device, characterized in that, include: The acquisition module is used to acquire pre-disaster and post-disaster images to be detected. A road blockage detection module is used to input the pre-disaster image and the post-disaster image into a target detection model, and obtain the change detection results of the road blockage area output by the target detection model; wherein, the target detection model includes an inter-guide neck network, and the inter-guide neck network includes: The temporal-specific enhancement module is used to apply edge structure enhancement to the pre-disaster local features of the pre-disaster image, apply spatial region enhancement to the post-disaster local features of the post-disaster image, and extract temporal difference priors based on the differences between the pre-disaster local features and the post-disaster local features. The bidirectional cross-guided fusion module is used to perform mutually constrained collaborative fusion of the shared global semantic features of the pre-disaster images and the post-disaster images with the temporal-specific fusion features, and output cross-guided fusion features; the temporal-specific fusion features are the fusion features of the enhanced pre-disaster local features, the enhanced post-disaster local features and the temporal difference prior. The spatial gating enhancement module is used to enhance the cross-guidance fusion features based on the enhanced pre-disaster local features, the enhanced post-disaster local features, and the spatial gating map generated by the temporal difference prior, and output the enhanced fusion features; the change detection result of the road blockage area is generated based on the enhanced fusion features.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the road obstruction detection method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the road obstruction detection method as described in any one of claims 1 to 7.

Images