A landslide-caused road-buried interactive target identification method, device, equipment and medium

Abstract

This invention provides a method, apparatus, device, and medium for identifying interactive targets caused by landslides burying roads. The method includes: acquiring remote sensing images of landslide-burial areas collected by unmanned aerial vehicles (UAVs); preprocessing the remote sensing images of the landslide-burial areas to obtain target area images; inputting the target area images into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps; performing intra-layer dynamic fusion of the multiple convolutional branch feature maps to obtain a fused image; performing a soft masking guided operation on the fused image to obtain a texture image; inputting the texture image into a target recognition model to identify landslide-burial areas, generating road burial anchor frames; and determining the identification result of the landslide-burial interactive target area based on the road burial anchor frames. This invention can improve the timeliness, accuracy, and automation of landslide-burial area identification.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method, apparatus, equipment and medium for identifying interactive targets buried by landslides. Background Technology

[0002] Landslides triggered by major disasters such as earthquakes often cause devastating damage to major transportation routes, especially in mountainous areas. Heavy rainfall and earthquakes frequently cause landslide slope instability, easily leading to complete physical disruption of the road network. In severe cases, this threatens lives and property and hinders earthquake relief efforts. Roads are the lifeline of earthquake relief; therefore, the rapid and accurate identification of landslide-damaged road areas is crucial for the deployment of emergency rescue forces, the planning of rescue routes, and subsequent resilient transportation reconstruction. It is a key link in ensuring the efficient conduct of disaster relief work.

[0003] Existing road damage monitoring technologies still have many problems that urgently need to be solved:

[0004] First, the spatial interaction between linear roads and planar landslides is not adequately characterized. Road burial and damage caused by landslides is essentially a dynamic, binary spatial interaction process, far more difficult to identify than the extraction of single ground features. Most current deep learning models focus on semantic segmentation or target monitoring tasks for single objects. Landslide technology emphasizes the morphological extraction of planar slopes, while road technology focuses on road morphological extraction and road crack identification. This perspective neglects the spatial "intersection" between landslides and roads—the damaged and buried area—and has not been the subject of in-depth technical research, lacking a targeted characterization of their interaction.

[0005] Secondly, multi-scale feature extraction is difficult. Traditional deep learning-based single-target monitoring models, lacking prior spatial context, struggle to accurately capture the relationship between landslide material and linear roads across multiple scales. They cannot adapt to the interaction characteristics of landslides and roads at different scales, significantly impacting recognition accuracy. Specifically, roads appear long and narrow at large scales but wide at small scales; landslides exhibit smaller morphology at large scales but larger morphology at small scales.

[0006] Third, the ability to suppress interference from complex backgrounds is weak. The textures of some damaged areas, such as gravel, collapsed buildings, and soil deposits, are extremely similar to natural bare land and farmland, making identification easily susceptible to interference. In addition, complex backgrounds can also lead to false detections and missed detections, making it difficult to meet the needs of accurate monitoring. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method, apparatus, equipment, and medium for identifying interactive targets buried by landslides. This can improve the timeliness, accuracy, and automation of identifying landslide-buried road areas.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] A method for identifying interactive targets buried by a landslide, comprising:

[0010] Acquire remote sensing images of road burial areas caused by landslides, collected by drones;

[0011] The remote sensing image of the road burial area is preprocessed to obtain the target area image;

[0012] The target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps.

[0013] The feature maps of the multiple convolutional branches are dynamically fused within the layer to obtain a fused image;

[0014] A soft mask-guided operation is performed on the fused image to obtain a texture image;

[0015] The texture image is input into the target recognition model to identify landslide-induced road burial areas, generating road burial anchor boxes. The target recognition model is trained according to the following process: a set of historical road burial areas remote sensing images is acquired, and semantic segmentation processing is performed on the historical road burial area remote sensing image set to obtain training segmentation images. After the training segmentation images are processed by soft masking guidance, they are input into the preset target recognition model to identify landslide-induced road burial areas through the road boundary compact loss function, and the training recognition result is obtained.

[0016] Based on the road burial anchor frame, the identification result of the interactive target area caused by the landslide to bury the road is determined.

[0017] Optionally, the remote sensing image of the road burial area is preprocessed to obtain an image of the target area, including:

[0018] The remote sensing image of the road buried area is denoised to obtain the first intermediate data;

[0019] The first intermediate data is normalized to obtain the second intermediate data;

[0020] The second intermediate data is subjected to color correction processing to obtain the target area image.

[0021] Optionally, the target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps, including:

[0022] The target region image is decoupled to obtain a convolution branch;

[0023] The convolutional branches are subjected to dilated convolution processing to obtain multiple convolutional branch feature maps.

[0024] Optionally, the feature maps of the multiple convolutional branches are dynamically fused within each layer to obtain a fused image, including:

[0025] Pooling is performed on the multiple convolutional branch feature maps to obtain a dynamic weight vector;

[0026] Based on the dynamic weight vector, the multiple convolutional branch feature maps are subjected to identity mapping to obtain multi-scale feature maps;

[0027] The multi-scale feature map is subjected to pixel semantic classification to obtain a fused image.

[0028] Optionally, based on the dynamic weight vector, an identity mapping is performed on the multiple convolutional branch feature maps to obtain multi-scale feature maps, including:

[0029] pass This yields multi-scale feature maps;

[0030] in, This represents a multi-scale feature map, where Id represents the identity mapping. This represents the output feature map of the i-th convolutional branch. Represents a dynamic weight vector. This represents element-wise multiplication.

[0031] Optionally, a soft mask-guided operation is performed on the fused image to obtain a texture image, including:

[0032] Based on the fused image, the background region in the target region image is obtained;

[0033] The background area is fused pixel by pixel to obtain a texture image.

[0034] Optionally, the target recognition model training process includes:

[0035] Acquire a set of remote sensing images of historical road burial areas; the set of remote sensing images of historical road burial areas includes ground truth boxes;

[0036] Using the initial identification model, road burial identification was performed on the remote sensing image set of historical road burial areas to obtain the initial road burial anchor frame;

[0037] Based on the compact loss function, the loss difference between the initial road burial anchor frame and the truth frame is obtained;

[0038] The initial recognition model is optimized using the loss difference to obtain the target recognition model.

[0039] Wherein, the compact loss function is:

[0040] ,

[0041] in, Represents the compact loss function. represents the amplitude control factor, k represents the curvature control factor, and z represents the offset. denoted as gradient magnitude control factor, and C represents a constant term.

[0042] Embodiments of the present invention also provide a device for identifying interactive targets buried by landslides, comprising:

[0043] The acquisition module is used to acquire remote sensing images of road burial areas caused by landslides, collected by drones.

[0044] The processing module is used to preprocess the remote sensing image of the road burial area to obtain a target area image; input the target area image into a semantic segmentation model for multi-layer convolutional segmentation to obtain multiple convolutional branch feature maps; perform intra-layer dynamic fusion on the multiple convolutional branch feature maps to obtain a fused image; perform a soft masking operation on the fused image to obtain a texture image; input the texture image into a target recognition model to identify the landslide-induced road burial area and generate a road burial anchor box; the target recognition model is trained according to the following process: acquire a set of historical road burial area remote sensing images, perform semantic segmentation on the historical road burial area remote sensing image set to obtain a training segmentation image, perform a soft masking operation on the training segmentation image, input it into a preset target recognition model to perform landslide-induced road burial area identification processing through a road boundary compact loss function to obtain a training recognition result; and determine the recognition result of the landslide-induced road burial interactive target area based on the road burial anchor box.

[0045] Embodiments of the present invention also provide a computing device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for identifying interactive targets buried by landslides according to the present invention.

[0046] Embodiments of the present invention also provide a computer-readable storage medium storing a program that, when executed by a processor, implements the method for identifying interactive targets buried by landslides as described in the present invention.

[0047] The above-described technical solution of the present invention has at least the following technical effects:

[0048] The method for identifying landslide-induced road burial interactive targets according to the present invention involves: acquiring remote sensing images of landslide-induced road burial areas collected by UAVs; preprocessing the remote sensing images of the road burial areas to obtain target area images; inputting the target area images into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps; performing intra-layer dynamic fusion on the multiple convolutional branch feature maps to obtain a fused image; performing a soft masking guidance operation on the fused image to obtain a texture image; inputting the texture image into a target recognition model to identify landslide-induced road burial areas and generating road burial anchor boxes; training the target recognition model according to the following process: acquiring a set of historical road burial area remote sensing images and performing semantic segmentation processing on the historical road burial area remote sensing image set to obtain training segmentation images; performing soft masking guidance processing on the training segmentation images and inputting them into a preset target recognition model to perform landslide-induced road burial area recognition processing through a road boundary compact loss function to obtain training recognition results; and determining the identification result of the landslide-induced road burial interactive target area based on the road burial anchor boxes. This will improve the timeliness, accuracy, and automation of identifying landslide-damaged road burial areas. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the method for identifying interactive targets buried by landslides according to the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the working principle of the method for identifying interactive targets buried by landslides in this invention;

[0051] Figure 3 This is a schematic diagram of the multi-layer convolutional segmentation process in the method for identifying interactive targets buried by landslides according to the present invention.

[0052] Figure 4 This is a schematic diagram of the intra-layer dynamic fusion of the method for identifying interactive targets buried by landslides in accordance with the present invention;

[0053] Figure 5 This is a schematic diagram of the identification device for interactive targets buried by landslides according to the present invention. Detailed Implementation

[0054] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0055] like Figure 1As shown, an embodiment of the present invention proposes a method for identifying interactive targets buried by landslides, comprising:

[0056] Step S1: Obtain remote sensing images of the road buried area caused by the landslide, collected by the drone;

[0057] Step S2: Preprocess the remote sensing image of the road burial area to obtain the target area image;

[0058] Step S3: Input the target region image into the semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps;

[0059] Step S4: Perform intra-layer dynamic fusion of the multiple convolutional branch feature maps to obtain a fused image;

[0060] Step S5: Perform a soft mask guidance operation on the fused image to obtain a texture image;

[0061] Step S6: Input the texture image into the target recognition model to identify the landslide-induced road burial area and generate road burial anchor boxes; The target recognition model is trained according to the following process: acquire a set of historical road burial area remote sensing images, perform semantic segmentation processing on the historical road burial area remote sensing image set to obtain training segmentation images, perform soft mask guidance processing on the training segmentation images, and input them into the preset target recognition model to perform landslide-induced road burial area identification processing through the road boundary compact loss function to obtain training recognition results;

[0062] Step S7: Based on the road burial anchor frame, determine the identification result of the landslide-induced road burial interactive target area.

[0063] In this embodiment, as Figure 1 As shown, in the method for identifying interactive targets buried by landslides, firstly, remote sensing images of the landslide-induced road burial area are acquired by a UAV; a UAV equipped with a multispectral imaging module is selected, with the flight altitude controlled at 50-100 meters to ensure coverage of the entire landslide-damaged road area, and the flight path adopts a grid-like traversal to avoid missing images; the image resolution is preferably set to 1080P or higher, and the acquisition is avoided during adverse weather conditions such as rain and direct sunlight to reduce image noise; high-quality, noise-free images of the landslide-damaged road area are acquired to provide a clean input data source for subsequent multi-scale segmentation and image fusion, avoiding segmentation errors and fusion distortions caused by noise in the original image;

[0064] like Figure 2As shown, the present invention preprocesses the remote sensing image of the road burial area to obtain a target area image; inputs the target area image into a semantic segmentation model for multi-layer convolutional segmentation to obtain multiple convolutional branch feature maps; performs intra-layer dynamic fusion on the multiple convolutional branch feature maps to obtain a fused image; performs a soft masking guidance operation on the fused image to obtain a texture image; inputs the texture image into a target recognition model to identify the landslide-induced road burial area and generates a road burial anchor box; the target recognition model is trained according to the following process: acquiring a set of historical road burial area remote sensing images, performing semantic segmentation on the historical road burial area remote sensing image set to obtain a training segmentation image, performing a soft masking guidance operation on the training segmentation image, and inputting it into a preset target recognition model to perform landslide-induced road burial area identification processing through a road boundary compact loss function to obtain a training recognition result; and determines the identification result of the landslide-induced road burial interactive target area based on the road burial anchor box. This method addresses the shortcomings of traditional methods, such as incomplete segmentation, significant background interference, and low recognition accuracy, and achieves accurate and efficient identification of interactive areas where roads are buried due to landslides.

[0065] In an optional embodiment of the present invention, step S2 involves preprocessing the remote sensing image of the road burial area to obtain a target area image, including:

[0066] Step S21: Denoise the remote sensing image of the road buried area to obtain the first intermediate data;

[0067] Step S22: Normalize the first intermediate data to obtain the second intermediate data;

[0068] Step S23: Perform color correction processing on the second intermediate data to obtain the target area image.

[0069] In this embodiment, the remote sensing image of the road buried area is first denoised to obtain the first intermediate data; the original image is pre-processed to lay the foundation for subsequent multi-scale segmentation and feature extraction. Specifically, the convolution kernel size is set to 3×3 and the standard deviation σ=0.5 during denoising. This parameter can effectively filter random noise in the image and suppress high-frequency interference while preserving key details such as road edges and landslide textures to the greatest extent, avoiding feature blurring caused by over-filtering.

[0070] In step S22, the first intermediate data is normalized to obtain the second intermediate data; the size normalization operation uniformly adjusts all images to the same pixel size, preferably 512×512 pixels, and uses bilinear interpolation to resample pixels, taking into account both processing efficiency and image smoothness, ensuring that the original images with different acquisition angles and resolutions have a unified input dimension, and avoiding model training deviation and recognition accuracy reduction caused by size differences.

[0071] In step S23, the second intermediate data is subjected to color correction processing to obtain the target area image. The color correction adopts a combination of grayscale stretching and white balance adjustment. First, the grayscale stretching is used to expand the dynamic range of grayscale in the image and enhance the grayscale contrast of the road and landslide areas. Then, the white balance adjustment is used to correct the color deviation under different lighting conditions, so that the texture features of the road and landslide in the image are clearly distinguishable, providing a high-quality image foundation for subsequent accurate segmentation of the road and landslide areas.

[0072] In an optional embodiment of the present invention, in step S3, the target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps, including:

[0073] Step S31: Decouple the target region image to obtain a convolution branch;

[0074] Step S32: Perform dilated convolution processing on the convolution branches to obtain multiple convolution branch feature maps.

[0075] In this embodiment, as Figure 3 As shown, in step S31, decoupling the target region image to obtain a convolution branch may include:

[0076] The target region image is used as the input feature map of the semantic segmentation model, denoted as X. The feature map has dimensions H×W×C, where H is the image height, W is the image width, and C is the number of channels.

[0077] Using large kernel convolution decomposition function Convolution with the original large kernel The decoupling is performed into n 3×3 small-parameter convolutional branches with different porosities;

[0078] The formula for solving the large kernel convolution integral is: ;

[0079] in, This represents the output feature map of the nth convolutional branch. Let X represent the kernel convolution solution function, and let X represent the input feature map. Indicates the original convolution size. Represents a sequence of void ratios;

[0080] Preferably, n=3, corresponding to 3 different inflation rates, to capture multi-scale features.

[0081] In step S32, the convolutional branch is subjected to dilated convolution processing to obtain the convolutional branch feature map;

[0082] Each branch performs a dilated convolution operation. , This represents the output feature map of the i-th convolutional branch. This represents a dilated convolution operation, where X represents the input feature map. This represents the void ratio of the i-th branch; three differentiated expansion ratios are set: small expansion ratio... Medium expansion rate High expansion rate It captures small road / landslide features, transitional scale features, and large-area road / landslide features respectively; it ensures that the number of parameters in each convolutional branch is only 1 / n of the original large kernel convolution, and the output feature dimension remains H×W×C, avoiding model overfitting caused by too many parameters.

[0083] In an optional embodiment of the present invention, step S4 involves performing intra-layer dynamic fusion of the plurality of convolutional branch feature maps to obtain a fused image, including:

[0084] Step S41: Pool the multiple convolutional branch feature maps to obtain a dynamic weight vector;

[0085] Step S42: Based on the dynamic weight vector, perform an identity mapping on the multiple convolutional branch feature maps to obtain multi-scale feature maps;

[0086] Step S43: Perform pixel semantic classification on the multi-scale feature map to obtain the fused image.

[0087] In this embodiment, step S41 involves pooling the multiple convolutional branch feature maps to obtain a dynamic weight vector, including:

[0088] Step S411: After concatenating the feature maps of the multiple convolutional branches, average pooling is performed to obtain global spatial features;

[0089] Step S412: Perform nonlinear processing on the global spatial features to obtain a dynamic weight vector.

[0090] In this embodiment, as Figure 4 As shown, in step S411, the multiple convolutional branch feature maps are concatenated and then average pooled to obtain global spatial features, which may include:

[0091] For the obtained n branch feature maps Serial operation through channel dimension ,in This indicates a channel concatenation operation, resulting in a concatenated feature map. The dimension becomes H×W×(n×C); for Perform Global Average Pooling (GAP) to aggregate spatial global information, using the following formula:

[0092] ,

[0093] in, This indicates a global average pooling operation. This represents the concatenated feature map, where H represents the height of the feature map and W represents the width of the feature map. express The pixel values ​​of all channels in the i-th row and j-th column; the output dimension is 1×1×(n×C);

[0094] In step S412, the global spatial features are subjected to nonlinear processing to obtain a dynamic weight vector, which may include:

[0095] First, dimensionality is reduced using 1×1 convolution, then activated by ReLU, followed by dimensionality increase using 1×1 convolution, and finally normalized using Softmax to generate a dynamic weight vector with a sum of 1. The formula is

[0096] ,

[0097] in, Represents a dynamic weight vector. This represents a 1×1 convolution operation. This represents the activation function. This indicates a normalization operation. This indicates a global average pooling operation. This represents the feature map after concatenation.

[0098] In step S42, based on the dynamic weight vector, the multiple convolutional branch feature maps are subjected to identity mapping to obtain multi-scale feature maps; an element fusion formula is used to achieve dynamic aggregation of multi-branch features. The element fusion formula is as follows:

[0099] ,

[0100] in, This represents a multi-scale feature map, where Id represents the identity mapping. This represents the output feature map of the i-th convolutional branch. Represents a dynamic weight vector. Indicates element-wise multiplication;

[0101] In step S43, pixel semantic classification is performed on the multi-scale feature map to obtain the fused image, which may include:

[0102] A 1×1 convolution is used to compress the number of channels in the multi-scale feature map to 2 channels, corresponding to the two semantic categories of "background" and "road-landslide interaction area". Then, the Softmax function is used to classify each pixel, resulting in the following expression:

[0103] ,

[0104] in, Indicates the classification result. This indicates a normalization operation. This represents a 1×1 convolution operation. This represents a multi-scale feature map, where p represents the kernel size and n represents the number of kernels.

[0105] Classification results Binarization is performed to obtain the fused image, and the expression is as follows:

[0106] ,

[0107] Where M represents the fused image, This represents the probability that a pixel (x, y) belongs to the road-landslide interaction region; pixels with a probability greater than 0.5 output by Softmax are classified as belonging to the road-landslide interaction region and marked as such. Pixels with a probability less than or equal to 0.5 are classified as background areas and marked as such. Finally, a fused image M of the road and the initial morphology of the landslide at multiple scales was obtained.

[0108] In an optional embodiment of the present invention, step S5, based on the fused image, involves performing a soft mask-guided operation on the fused image to obtain a texture image, including:

[0109] Step S51: Obtain the background region in the target region image based on the fused image;

[0110] Step S52: Perform pixel-by-pixel fusion on the background area to obtain a texture image.

[0111] In this embodiment, for the fused image M, where This indicates that the pixel represents the foreground of the road and landslide interaction, and is a key area for identification. This indicates that the pixel is part of the background area;

[0112] In this embodiment, step S51, obtaining the background region in the target region image based on the fused image, may include:

[0113] The classification results of pixel semantic classification on multi-scale feature maps are binarized. When pixel (x, y) satisfies the classification result... When the pixel is determined not to belong to the road-landslide interaction area, it is marked as... All marked as The set of pixels is the desired background area.

[0114] In step S52, the background region is fused pixel by pixel to obtain a texture image, which may include:

[0115] Define a texture image by fusing the original multispectral image and the fused image M pixel by pixel, using the following formula:

[0116] ,

[0117] in, Represents a texture image. M represents the target region image, and M represents the fused image. Indicates the brightness suppression coefficient. Indicates the base brightness offset;

[0118] The fused texture image is obtained by calculating pixel by pixel using the formula described above. The image retains the complete texture of the foreground area (road, landslide) while darkening the background area to reduce background interference.

[0119] In an optional embodiment of the present invention, step S6, in which the target recognition model is trained according to the following process, may include:

[0120] Step S61: Obtain a set of remote sensing images of historical road burial areas; the set of remote sensing images of historical road burial areas includes ground truth boxes;

[0121] Step S62: Perform semantic segmentation processing on the remote sensing image set of the historical road burial area to obtain training segmentation images;

[0122] Step S63: After performing soft mask-guided processing on the training segmentation image, input it into the initial recognition model to perform road burial recognition on the remote sensing image set of historical road burial areas, and obtain the initial road burial anchor frame.

[0123] Step S64: Based on the compact loss function, obtain the loss difference between the initial road burial anchor frame and the truth frame;

[0124] Step S65: Optimize the hyperparameters of the initial recognition model based on the loss difference to obtain the target recognition model.

[0125] In this embodiment, the initial road burial anchor frame is represented by a feature vector (x, y, w, h), where x and y are the coordinates of the anchor frame's center point, and w and h are the width and height of the anchor frame; the offset between the anchor frame and the truth box is defined. The formula for calculating the horizontal axis offset is:

[0126] ,

[0127] in, This represents the horizontal axis offset. This represents the x-coordinate of the center point of the truth box. This represents the x-coordinate of the center point of the initial road burial anchor frame. This indicates the width of the center point of the initial road burial anchor frame;

[0128] The formula for calculating the ordinate offset is:

[0129] ,

[0130] in, This indicates the offset of the vertical axis. This represents the y-coordinate of the center point of the truth box. This represents the ordinate of the center point of the initial road burial anchor frame. This indicates the height of the center point of the initial road burial anchor frame;

[0131] The formula for calculating the width offset is:

[0132] ,

[0133] in, Indicates the width offset. Indicates the width of the truth box; This indicates the width of the center point of the initial road burial anchor frame;

[0134] The formula for calculating the height offset is:

[0135] ,

[0136] in, Indicates the height offset. Indicates the height of the truth box; This indicates the height of the center point of the initial road burial anchor frame;

[0137] Constitute the complete offset This is used to correct the anchor box position so that the corrected anchor box is aligned with the truth box;

[0138] In the calculation of anchor frame correction loss, |z|<1 (where z is the offset). Anchor boxes with |z|≥1 are called compact samples, which are high-quality anchor boxes that accurately cover the interaction region. Anchor boxes with |z|≥1 are called relaxed samples, which are low-quality anchor boxes that deviate from the interaction region. A nonlinear gradient function is designed to suppress excessive gradients of low-quality anchor boxes and enhance the weak gradients of high-quality anchor boxes. The gradient formula is:

[0139] ,

[0140] in, Represents the compact loss function. represents the amplitude control factor, k represents the curvature control factor, and z represents the offset. This represents the gradient magnitude control factor, and sign represents the sign function.

[0141] Integrating the gradient formula yields the compact loss function:

[0142] ,

[0143] in, Represents the compact loss function. represents the amplitude control factor, k represents the curvature control factor, and z represents the offset. This represents the gradient magnitude control factor, and C represents the constant term.

[0144] The compact loss function is used for backpropagation optimization of the model. By minimizing the compact loss function, the anchor box position is continuously corrected, and finally the identification result of the interactive target of landslide-induced road burial is obtained, accurately marking the location and range of the road burial area.

[0145] In an optional embodiment of the present invention, step S6, inputting the texture image into a target recognition model to identify landslide-induced road burial areas and generating road burial anchor frames, may include:

[0146] Step S66: Input the texture image into the target recognition model that minimizes the compact loss function to perform landslide recognition and road recognition, and generate road burial anchor frames for the landslide-induced road burial area.

[0147] In this embodiment, a target recognition model that minimizes the compact loss function is used to identify landslides and roads, ultimately obtaining accurate identification results of the interactive target of landslide-induced road burial. The road burial anchor frame can accurately mark the location and range of the road burial area.

[0148] In an optional embodiment of the present invention, step S7, determining the identification result of the landslide-induced road burial interactive target area based on the road burial anchor frame, may include:

[0149] Step S71: Overlay the road burial anchor frame onto the fused image, simultaneously display the foreground and background areas, and identify the road and the landslide interacting with the road in the interaction area where the road burial anchor frame is located as the identification result of the landslide-induced road burial interaction target area, providing a rapid location reference for the burial area for emergency monitoring.

[0150] The present invention provides a precise fused image for soft mask fusion through multi-scale segmentation, ensuring that the fused image can accurately distinguish between the foreground and background, providing high-quality input for subsequent detection; it reduces background interference and highlights the interactive foreground through soft mask fusion, providing a clear anchor box detection scenario for compact loss function optimization and reducing the generation of low-quality anchor boxes; it solves the gradient imbalance problem through compact loss function optimization, further improving the anchor box positioning accuracy and compensating for minor errors in previous steps; by forming a cascaded architecture of "segmentation-fusion-optimization", it completely solves the pain points of traditional methods such as "incomplete segmentation, large background interference, and low recognition accuracy", achieving accurate and efficient identification of interactive areas buried by landslides.

[0151] The solution of this invention does not require manpower, professional interpreters, or fine preprocessing of remote sensing images when identifying interactive targets buried by landslides. It is highly efficient. At the same time, the cascaded framework provides ideas for subsequent interactive target identification and offers a good solution for multi-scale segmentation and suppression of complex background interference.

[0152] like Figure 5 As shown, embodiments of the present invention also provide a landslide-induced road burial identification device 50, comprising:

[0153] The acquisition module 51 is used to acquire remote sensing images of the road buried area caused by landslides collected by the UAV;

[0154] Processing module 52 is used to preprocess the remote sensing image of the road burial area to obtain a target area image; input the target area image into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps; perform intra-layer dynamic fusion on the multiple convolutional branch feature maps to obtain a fused image; perform a soft masking guidance operation on the fused image to obtain a texture image; input the texture image into a target recognition model to identify the landslide-induced road burial area and generate a road burial anchor box; the target recognition model is trained according to the following process: acquiring a set of historical road burial area remote sensing images, performing semantic segmentation processing on the historical road burial area remote sensing image set to obtain a training segmentation image, performing a soft masking guidance operation on the training segmentation image, and inputting it into a preset target recognition model to perform landslide-induced road burial area identification processing through a road boundary compact loss function to obtain a training recognition result; and determining the recognition result of the landslide-induced road burial interactive target area based on the road burial anchor box.

[0155] Optionally, the remote sensing image of the road burial area is preprocessed to obtain an image of the target area, including:

[0156] The remote sensing image of the road buried area is denoised to obtain the first intermediate data;

[0157] The first intermediate data is normalized to obtain the second intermediate data;

[0158] The second intermediate data is subjected to color correction processing to obtain the target area image.

[0159] Optionally, the target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps, including:

[0160] The target region image is decoupled to obtain a convolution branch;

[0161] The convolutional branches are subjected to dilated convolution processing to obtain multiple convolutional branch feature maps.

[0162] Optionally, the feature maps of the multiple convolutional branches are dynamically fused within each layer to obtain a fused image, including:

[0163] Pooling is performed on the multiple convolutional branch feature maps to obtain a dynamic weight vector;

[0164] Based on the dynamic weight vector, the multiple convolutional branch feature maps are subjected to identity mapping to obtain multi-scale feature maps;

[0165] The multi-scale feature map is subjected to pixel semantic classification to obtain a fused image.

[0166] Optionally, based on the dynamic weight vector, an identity mapping is performed on the multiple convolutional branch feature maps to obtain multi-scale feature maps, including:

[0167] pass This yields multi-scale feature maps;

[0168] in, This represents a multi-scale feature map, where Id represents the identity mapping. This represents the output feature map of the i-th convolutional branch. Represents a dynamic weight vector. This represents element-wise multiplication.

[0169] Optionally, a soft mask-guided operation is performed on the fused image to obtain a texture image, including:

[0170] Based on the fused image, the background region in the target region image is obtained;

[0171] The background area is fused pixel by pixel to obtain a texture image.

[0172] Optionally, the target recognition model training process includes:

[0173] Acquire a set of remote sensing images of historical road burial areas; the set of remote sensing images of historical road burial areas includes ground truth boxes;

[0174] Semantic segmentation processing was performed on a set of remote sensing images of historical road burial areas to obtain training segmented images;

[0175] After performing soft mask-guided processing on the training segmented image, it is input into the initial recognition model to perform road burial recognition on the remote sensing image set of historical road burial areas, and obtain the initial road burial anchor frame.

[0176] Based on the compact loss function, the loss difference between the initial road burial anchor frame and the truth frame is obtained;

[0177] The initial recognition model is optimized using the loss difference to obtain the target recognition model; wherein, the compact loss function is:

[0178] ,

[0179] in, Represents the compact loss function. represents the amplitude control factor, k represents the curvature control factor, and z represents the offset. denoted as gradient magnitude control factor, and C represents a constant term.

[0180] It should be noted that all implementation methods in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effect.

[0181] Embodiments of the present invention also provide a computing device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for identifying interactive targets buried by landslides according to the present invention. All implementations in the above method embodiments are applicable to the embodiments of this computing device and can achieve the same technical effects.

[0182] Embodiments of the present invention also provide a computer-readable storage medium storing a program that, when executed by a processor, implements the method for identifying interactive targets buried by landslides as described in this invention. All implementations in the above method embodiments are applicable to embodiments of this computing device and can achieve the same technical effects.

[0183] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0184] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0185] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0186] 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0187] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0188] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion 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 of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0189] Furthermore, it should be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Moreover, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order; some steps can be executed in parallel or independently of each other. Those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in hardware, firmware, software, or a combination thereof. This is something that those skilled in the art can achieve by using their basic programming skills after reading the description of the present invention.

[0190] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing device. The computing device can be a known general-purpose device. Therefore, the object of the present invention can also be achieved simply by providing a program product containing program code for implementing the method or apparatus. That is, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any known storage medium or any storage medium developed in the future. It should also be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent to the present invention. Furthermore, the steps for performing the above series of processes can naturally be performed in the order described, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel or independently of each other.

[0191] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying landslide-induced road burial interaction targets, characterized in that, include: Acquire remote sensing images of the road buried by landslides, collected by drones; The remote sensing image of the road burial area is preprocessed to obtain the target area image; The target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing to obtain multiple convolutional branch feature maps. The feature maps of the multiple convolutional branches are dynamically fused within the layer to obtain a fused image; A soft mask-guided operation is performed on the fused image to obtain a texture image; The texture image is input into the target recognition model to identify the landslide-induced road burial area and generate a road burial anchor frame. The target recognition model is trained according to the following process: acquiring a set of remote sensing images of historical road buried areas, performing semantic segmentation on the set of remote sensing images of historical road buried areas to obtain training segmented images, performing soft masking guided operation on the training segmented images, and inputting them into a preset target recognition model to perform landslide-induced road buried areas recognition processing through a road boundary compact loss function to obtain training recognition results. Based on the road burial anchor frame, the identification result of the interactive target area caused by the landslide to bury the road is determined.

2. The method of claim 1, wherein, The remote sensing image of the road burial area is preprocessed to obtain the target area image, including: The remote sensing image of the road buried area is denoised to obtain the first intermediate data; The first intermediate data is normalized to obtain the second intermediate data; The second intermediate data is subjected to color correction processing to obtain the target area image.

3. The method for identifying interactive targets buried by landslides as described in claim 1, characterized in that, The target region image is input into a semantic segmentation model for multi-layer convolutional segmentation processing, resulting in multiple convolutional branch feature maps, including: The target region image is decoupled to obtain a convolution branch; The convolutional branches are subjected to dilated convolution processing to obtain multiple convolutional branch feature maps.

4. The method for identifying interactive targets buried by landslides as described in claim 1, characterized in that, The multiple convolutional branch feature maps are dynamically fused within a layer to obtain a fused image, including: Pooling is performed on the multiple convolutional branch feature maps to obtain a dynamic weight vector; Based on the dynamic weight vector, the multiple convolutional branch feature maps are subjected to identity mapping to obtain multi-scale feature maps; The multi-scale feature map is subjected to pixel semantic classification to obtain a fused image.

5. The method for identifying interactive targets buried by landslides as described in claim 4, characterized in that, Based on the dynamic weight vector, an identity mapping is performed on the multiple convolutional branch feature maps to obtain multi-scale feature maps, including: pass This yields multi-scale feature maps; in, This represents a multi-scale feature map, where Id represents the identity mapping. This represents the output feature map of the i-th convolutional branch. Represents a dynamic weight vector. This represents element-wise multiplication.

6. The method for identifying interactive targets buried by landslides as described in claim 1, characterized in that, A soft mask-guided operation is performed on the fused image to obtain a texture image, including: Based on the fused image, the background region in the target region image is obtained; The background area is fused pixel by pixel to obtain a texture image.

7. The method for identifying interactive targets buried by landslides as described in claim 1, characterized in that, The target recognition model training process includes: Acquire a set of remote sensing images of historical road burial areas; the set of remote sensing images of historical road burial areas includes ground truth boxes; Semantic segmentation processing was performed on a set of remote sensing images of historical road burial areas to obtain training segmented images; After performing soft mask-guided processing on the training segmented image, it is input into the initial recognition model to perform road burial recognition on the remote sensing image set of historical road burial areas, and obtain the initial road burial anchor frame. Based on the compact loss function, the loss difference between the initial road burial anchor frame and the truth frame is obtained; The initial recognition model is optimized using the loss difference to obtain the target recognition model. Wherein, the compact loss function is: , in, Represents the compact loss function. represents the amplitude control factor, k represents the curvature control factor, and z represents the offset. denoted as gradient magnitude control factor, and C represents a constant term.

8. A device for identifying interactive targets buried by landslides, characterized in that, include: The acquisition module is used to acquire remote sensing images of road burial areas caused by landslides, collected by drones. The processing module is used to preprocess the remote sensing image of the road buried area to obtain a target area image; input the target area image into a semantic segmentation model for multi-layer convolutional segmentation to obtain multiple convolutional branch feature maps; perform intra-layer dynamic fusion on the multiple convolutional branch feature maps to obtain a fused image; perform a soft masking operation on the fused image to obtain a texture image; input the texture image into a target recognition model to identify the landslide-induced road buried area and generate road buried anchor boxes. The target recognition model is trained according to the following process: acquiring a set of remote sensing images of historical road burial areas, performing semantic segmentation on the set of remote sensing images of historical road burial areas to obtain training segmented images, performing soft masking guidance on the training segmented images, inputting them into a preset target recognition model to perform landslide-induced road burial area recognition processing through a road boundary compact loss function to obtain training recognition results; and determining the recognition result of the landslide-induced road burial interactive target area based on the road burial anchor frame.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images