A deep learning-based slope disaster light perception system, device and platform

Abstract

The present application relates to the technical field of geological disaster monitoring, and specifically discloses a slope disaster nimble sensing system, device and platform based on deep learning, which comprises a slope disaster sample library construction module, a slope disaster information real-time extraction model construction module, a global feature extraction module, a non-local feature extraction module, a slope disaster information real-time extraction model training module and a quantization compression module. The system can realize model quantization deployment on an embedded device, and the device is carried on a UAV platform. Image data is acquired through the camera module of the UAV, and is transmitted to the platform in real time. The platform extracts slope information from the image data, and transmits the information to the ground control terminal in real time, thereby solving the problems of slow speed and low accuracy in acquiring slope information during post-disaster rescue, and realizing real-time sensing of slope disasters in the post-disaster rescue area.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, specifically to a lightweight slope disaster sensing system, equipment, and platform based on deep learning. Background Technology

[0002] Earthquake-induced landslides are characterized by their large number, scale, and simultaneous occurrence in multiple locations. Rapidly and accurately identifying landslide boundaries and estimating the affected area is crucial for landslide emergency rescue. With the widespread use of drones in post-landslide rescue efforts, multi-view, ultra-high-resolution drone imagery presents new challenges for landslide extraction. The difficulty in distinguishing landslides from exposed ground exacerbates the problem. Furthermore, current traditional landslide extraction methods involve a two-step process: first generating a DOM (Orthophoto Map) from the original drone imagery, then extracting the landslide. Generating orthophotos relies on high-performance computing equipment, is time-consuming, and fails to provide timely and accurate data support for rescue and road clearing. Therefore, there is an urgent need to develop a precise, real-time landslide extraction method.

[0003] In recent years, with the development of deep neural network technology, the technique of using convolutional neural networks to extract landslide disaster information from images has gradually matured. However, when processing high-resolution UAV imagery, models with a large number of parameters are required, making them difficult to deploy on embedded platforms and affecting the real-time performance of landslide extraction. At the same time, using smaller models leads to reduced accuracy in landslide extraction, posing a challenge to the precision of landslide-based emergency rescue. Therefore, it is necessary to design a precise and real-time landslide extraction method and system to meet the specific needs of landslide disaster information extraction from UAV imagery, providing accurate data support for emergency rescue and road clearing. This method requires deploying a quantized and compressed real-time extraction model on an embedded platform or device to achieve efficient extraction of landslide disaster information. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a lightweight slope disaster perception system, device, and platform based on deep learning, which enables accurate and real-time perception of slope disasters, solves the problem of lack of landslide interpretation data to support decision-making during the golden rescue period after a disaster, and resolves the problems mentioned in the background art above.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a lightweight slope disaster perception system based on deep learning, the system comprising:

[0006] Slope disaster sample database construction module: Collect UAV image data of the disaster area, label the slope disaster information in it, and then form a slope disaster sample database after data preprocessing;

[0007] Real-time slope disaster information extraction model construction module: used to construct a real-time slope disaster information extraction model;

[0008] Global feature extraction module and non-local feature extraction module: used to construct the global feature extraction module and non-local feature extraction module to obtain more comprehensive contextual semantic information. The global feature extraction module and non-local feature extraction module are connected in parallel with the backbone network of the real-time slope disaster information extraction model as an extension of the backbone network.

[0009] Real-time slope disaster information extraction model training module: The real-time slope disaster information extraction model is trained using a slope disaster sample library, and the model parameters are adjusted.

[0010] The quantization and compression module performs quantization and compression on the real-time extraction model of slope disaster information. Guided by the minimum relative entropy, it nonlinearly maps the model's weight data type FP32 to INT8 and performs quantization and compression on the model by calculating the pruning threshold and scale factor.

[0011] Preferably, the backbone network structure of the real-time slope disaster information extraction model has 101 layers, including 4 convolutional combinations, one global average pooling layer, one fully connected layer and one softmax layer. Each convolutional combination includes 3, 4, 23 and 3 residual modules in sequence, and each residual module contains 3 convolutional layers and 1 regularization layer.

[0012] Preferably, the data preprocessing includes rotation, scaling, random cropping, and color transformation.

[0013] Preferably, the global feature extraction module includes a global average pooling layer and two fully connected layers, comprising pooling operations, convolution operations, ReLU activation function, and sigmoid activation function; the specific calculations of this module include the following:

[0014] First, global average pooling is used to aggregate the feature values ​​for each channel. This operation adds up all feature values ​​and calculates the average, as shown in the following formula:

[0015]

[0016] In the formula: H represents the statistical value obtained after global average pooling of the c-th channel of the feature map; H represents the height of the feature map; W represents the width of the feature map. This represents the c-th channel of the feature map;

[0017] Secondly, the relationship between different channels is captured using two different fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the sigmoid activation function. The calculation formula is shown below:

[0018]

[0019] In the formula: s is the weight parameter obtained from learning different channels; Represents the sigmoid activation function; This represents the weight parameters of the first fully connected layer; This represents the weight parameters of the second fully connected layer;

[0020] Finally, the activation values ​​of each channel learned by the global feature extraction module are multiplied by the original features at the corresponding positions on the input feature map, as shown in the following formula:

[0021]

[0022] In the formula: This represents the new feature map with dimensions c×h×w obtained after processing by the global feature extraction module.

[0023] Preferably, the non-local feature extraction module includes a spatial attention module and a channel attention mechanism module, which consists of convolution operations, a softmax function, a reshape function, and a transpose function; the specific calculations of this module include the following:

[0024] By combining the input feature map with an attention mechanism, the attention matrix S is calculated. This matrix expresses the weight of attention at each location on other locations in the feature map. The formula for calculating the attention matrix S is as follows:

[0025]

[0026] In the formula: B and C are the new feature maps of size c×h×w obtained after the original feature maps are convolved; The first of the S matrix j Line number i The characteristic values ​​of the column;

[0027] After obtaining the attention matrix S, the spatial attention feature map P is calculated using the following formula:

[0028]

[0029] In the formula: Let be the feature value at the j-th position of feature map P; D is the feature map of size c×h×w obtained after the original feature map is convolved. Let the feature value at the j-th position of the input feature map A be denoted as ; Parameters used to control the weights of eigenvalues;

[0030] The channel attention matrix X is calculated by combining the input feature map with the attention mechanism. The X matrix expresses the attention magnitude of each channel to other channels, and its calculation formula is as follows:

[0031]

[0032] In the formula: Let the feature value be the feature value at the i-th position of the input feature map A; Let X be the eigenvalue of the j-th row and i-th column of matrix X;

[0033] After obtaining the attention matrix X, the channel attention feature map E can be calculated using the following formula:

[0034]

[0035] In the formula: Let J be the feature value at the j-th position of feature map E; Parameters used to control the weights of eigenvalues;

[0036] After obtaining the spatial attention feature map and the channel attention feature map, the two feature maps are added element by element to obtain a new set of feature maps with dimensions c×h×w.

[0037] Preferably, the quantization compression in the quantization compression module includes the following steps:

[0038] First, determine the pruning threshold for model quantization and remove outliers. The formula for calculating the pruning threshold is as follows:

[0039]

[0040] In the formula: These are the weight parameters for the FP32 type model; This is the clipping threshold;

[0041] Then, the quantization scaling factor is calculated by using a pruning threshold. The floating-point dynamic range is divided into 256 intervals. The formula for calculating the scaling factor is as follows:

[0042]

[0043] In the formula: Scale factor;

[0044] Finally, guided by the principle of minimizing relative entropy, the model is quantized using a pruning threshold and a scaling factor, and the model's weight data type FP32 is non-linearly mapped to INT8, as shown in the following formula:

[0045]

[0046] In the formula: These are model weight parameters of type INT8.

[0047] In addition, to achieve the above objectives, the present invention also provides the following technical solution: a device, the device including at least one processor; and at least one processor is equipped with the aforementioned lightweight slope disaster sensing system.

[0048] In addition, to achieve the above objectives, the present invention also provides the following technical solution: a platform, the platform including the aforementioned device, camera module, positioning module, power supply module, and communication module;

[0049] Camera module: Used to acquire remote sensing images of the ground during the drone's flight;

[0050] Positioning module: Used to obtain the current position of the drone, add position information to the drone imagery, and improve the accuracy of the image data;

[0051] Communication module: Used to receive flight control signals and camera control signals, and transmit slope information extraction results and images captured by the camera module back to the UAV ground control terminal;

[0052] Power supply module: Used to supply power to the equipment, camera module, positioning module, and communication module.

[0053] The beneficial effects of this invention are: the system of this invention can realize the quantitative deployment of models on embedded devices, mount the device on the UAV platform, acquire image data through the camera module of the UAV and transmit it to the platform in real time, extract slope information from the image data through the platform and transmit it to the ground control terminal in real time, thereby solving the problem of slow speed and low accuracy of acquiring slope information during disaster relief, and realizing real-time perception of slope disasters in the disaster relief area. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the system components of the present invention;

[0055] Figure 2 This is a schematic diagram of the global feature extraction module of the present invention;

[0056] Figure 3 This is a schematic diagram of the nonlocal feature extraction module of the present invention;

[0057] Figure 4 This is a schematic diagram illustrating the quantization compression principle of the present invention;

[0058] Figure 5 This is a schematic diagram of the platform architecture of the present invention;

[0059] Among them, 110-Slope disaster sample database construction module; 120-Slope disaster information real-time extraction model construction module; 130-Global feature extraction module and non-local feature extraction module; 140-Slope disaster information real-time extraction model training module; 150-Quantization compression module. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1-5 This invention provides a technical solution: a lightweight slope disaster perception system based on deep learning, such as... Figure 1 As shown, the system includes:

[0062] Module 110 for constructing a slope disaster sample database: collects UAV image data of the disaster-stricken area, annotates the slope disaster information in the data, and then forms a slope disaster sample database after data preprocessing.

[0063] Further data preprocessing includes rotation, scaling, random cropping, and color transformation.

[0064] Module 120 for constructing a real-time slope disaster information extraction model: This module is used to construct a real-time slope disaster information extraction model.

[0065] The backbone network structure of the real-time slope disaster information extraction model has 101 layers, including 4 convolutional combinations, one global average pooling layer, one fully connected layer and one softmax layer. Each convolutional combination includes 3, 4, 23 and 3 residual modules respectively. Each residual module contains 3 convolutional layers and 1 regularization layer.

[0066] Global feature extraction module and non-local feature extraction module 130: used to construct the global feature extraction module and non-local feature extraction module to obtain more comprehensive contextual semantic information. The global feature extraction module and non-local feature extraction module are connected in parallel with the backbone network of the real-time slope disaster information extraction model as an extension of the backbone network.

[0067] The global feature extraction module includes a global average pooling layer and two fully connected layers, comprising pooling operations, convolution operations, ReLU activation function, and sigmoid activation function. The specific computations of this module are as follows:

[0068] First, global average pooling is used to aggregate the feature values ​​for each channel. This operation adds up all feature values ​​and calculates the average, as shown in the following formula:

[0069]

[0070] In the formula: H represents the statistical value obtained after global average pooling of the c-th channel of the feature map; H represents the height of the feature map; W represents the width of the feature map. This represents the c-th channel of the feature map;

[0071] Secondly, the relationship between different channels is captured using two different fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the sigmoid activation function. The calculation formula is shown below:

[0072]

[0073] In the formula: s is the weight parameter obtained from learning different channels; Represents the sigmoid activation function; This represents the weight parameters of the first fully connected layer; This represents the weight parameters of the second fully connected layer;

[0074] Finally, the activation values ​​of each channel learned by the global feature extraction module are multiplied by the original features at the corresponding positions on the input feature map, as shown in the following formula:

[0075]

[0076] In the formula: This represents the new feature map with dimensions c×h×w obtained after processing by the global feature extraction module.

[0077] Furthermore, the non-local feature extraction module includes a spatial attention module and a channel attention mechanism module, which consists of convolution operations, a softmax function, a reshape function, and a transpose function; the specific calculations of this module include the following:

[0078] By combining the input feature map with an attention mechanism, the attention matrix S is calculated. This matrix expresses the weight of attention at each location on other locations in the feature map. The formula for calculating the attention matrix S is as follows:

[0079]

[0080] In the formula: B and C are the new feature maps of size c×h×w obtained after the original feature maps are convolved; The first of the S matrix j Line number i The characteristic values ​​of the column;

[0081] After obtaining the attention matrix S, the spatial attention feature map P is calculated using the following formula:

[0082]

[0083] In the formula: Let be the feature value at the j-th position of feature map P; D is the feature map of size c×h×w obtained after the original feature map is convolved. Let the feature value at the j-th position of the input feature map A be denoted as ; Parameters used to control the weights of eigenvalues;

[0084] The channel attention matrix X is calculated by combining the input feature map with the attention mechanism. The X matrix expresses the attention magnitude of each channel to other channels, and its calculation formula is as follows:

[0085]

[0086] In the formula: Let the feature value be the feature value at the i-th position of the input feature map A; Let X be the eigenvalue of the j-th row and i-th column of matrix X;

[0087] After obtaining the attention matrix X, the channel attention feature map E can be calculated using the following formula:

[0088]

[0089] In the formula: Let J be the feature value at the j-th position of feature map E; Parameters used to control the weights of eigenvalues;

[0090] After obtaining the spatial attention feature map and the channel attention feature map, the two feature maps are added element by element to obtain a new set of feature maps with dimensions c×h×w.

[0091] Module 140 for training the real-time slope disaster information extraction model: Uses a slope disaster sample library to train the real-time slope disaster information extraction model and adjusts the model parameters.

[0092] The quantization and compression module 150 performs quantization and compression on the real-time extraction model of slope disaster information. Guided by the minimum relative entropy, it nonlinearly maps the model's weight data type FP32 to INT8 and performs quantization and compression on the model by calculating the pruning threshold and scale factor.

[0093] Furthermore, the quantization compression includes the following steps:

[0094] First, determine the pruning threshold for model quantization and remove outliers. The formula for calculating the pruning threshold is as follows:

[0095]

[0096] In the formula: These are the weight parameters for the FP32 type model; This is the clipping threshold;

[0097] Then, the quantization scaling factor is calculated by using a pruning threshold. The floating-point dynamic range is divided into 256 intervals. The formula for calculating the scaling factor is as follows:

[0098]

[0099] In the formula: Scale factor;

[0100] Finally, guided by the principle of minimizing relative entropy, the model is quantized using a pruning threshold and a scaling factor, and the model's weight data type FP32 is non-linearly mapped to INT8, as shown in the following formula:

[0101]

[0102] In the formula: These are model weight parameters of type INT8.

[0103] In one specific embodiment, a sensing method utilizing the aforementioned sensing system, or a landslide extraction method that is both accurate and real-time, comprises the following steps:

[0104] Step 1: Collect UAV imagery data of the disaster-stricken area, label the slope disaster information, and form a sample database dataset.

[0105] The specific steps of the slope disaster information annotation method are as follows: SFM technology is used to estimate the camera pose and the 3D point cloud of the scene for each image. Based on the camera pose and point cloud, perspective views of multiple images are generated. Projecting the perspective views onto a plane yields the position and color information of each pixel. According to the camera's intrinsic and extrinsic parameters, the projected pixels are corrected and translated to obtain an orthophoto unaffected by terrain and shooting angle. Processing the orthophoto yields a high-resolution DOM.

[0106] The slope hazard information on multiple DOMs was annotated using a visual interpretation method, and the annotated DOMs and slope hazard information samples were stored in a sample database.

[0107] Step 2: Construct a real-time slope disaster information extraction model. The model captures long-range contextual semantic information by expanding the receptive field of the network, while preserving the edge structure of the landslide. The backbone network structure consists of 101 layers, with the main structure including four convolutional combinations. Each convolutional combination successively includes 3, 4, 23, and 3 residual modules. Each residual module contains 3 convolutional layers and 1 regularization layer. In addition, the backbone network also includes a global average pooling layer, a fully connected layer, and a softmax layer.

[0108] The specific steps for extracting the output feature map of UAV remote sensing images using the backbone network for real-time extraction of slope disaster information are as follows:

[0109] After inputting UAV remote sensing images into the slope disaster information real-time extraction backbone network, the images sequentially pass through one convolutional layer and one pooling layer, followed by four convolutional combinations. Each convolutional combination downsamples the feature map by a factor of 2, ultimately yielding an image with the same size as the input image. The output feature map.

[0110] Step 3: Construct a global feature extraction module and a non-local feature extraction module to obtain more comprehensive contextual semantic information; connect the global and non-local feature extraction modules in parallel with the real-time slope disaster information extraction backbone network as an extension of the backbone network to improve the feature extraction effect; input the feature maps output by the backbone network into the global and non-local feature extraction modules respectively to obtain two new feature maps; upsample these two feature maps by a factor of 16, and then concatenate the upsampled feature maps to form a new feature map. Perform a 1×1 convolution operation on the new feature map to transform it into a single-channel probability map. Set a threshold of 0.5, segment the parts with a probability greater than 0.5 into foreground, and segment the remaining parts into background.

[0111] like Figure 2 As shown, the global feature extraction module includes a global average pooling layer and two fully connected layers, which include pooling operations, convolution operations, ReLU activation function, and sigmoid activation function; the calculation process includes feature matrix multiplication and matrix average calculation.

[0112] The specific calculations in this module include the following:

[0113] First, global average pooling is used to aggregate the feature values ​​for each channel. This operation adds up all feature values ​​and calculates the average, as shown in the following formula:

[0114]

[0115] In the formula: H represents the statistical value obtained after global average pooling of the c-th channel of the feature map; H represents the height of the feature map; W represents the width of the feature map. This represents the c-th channel of the feature map;

[0116] Secondly, the relationship between different channels is captured using two different fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the sigmoid activation function. The calculation formula is shown below:

[0117]

[0118] In the formula: s is the weight parameter obtained from learning different channels; Represents the sigmoid activation function; This represents the weight parameters of the first fully connected layer; This represents the weight parameters of the second fully connected layer;

[0119] Finally, the activation values ​​of each channel learned by the global feature extraction module are multiplied by the original features at the corresponding positions on the input feature map, as shown in the following formula:

[0120]

[0121] In the formula: This represents the new feature map with dimensions c×h×w obtained after processing by the global feature extraction module.

[0122] like Figure 3 As shown, the non-local feature extraction module includes a spatial attention module and a channel attention mechanism module, which consists of convolution operations, a softmax function, a reshape function, and a transpose function. The calculation process includes feature matrix transformation, feature matrix multiplication, and feature matrix addition. The specific calculations of this module are as follows:

[0123] By combining the input feature map with an attention mechanism, the attention matrix S is calculated. This matrix expresses the weight of attention at each location on other locations in the feature map. The formula for calculating the attention matrix S is as follows:

[0124]

[0125] In the formula: B and C are the new feature maps of size c×h×w obtained after the original feature maps are convolved; The first of the S matrix j Line number i The characteristic values ​​of the column;

[0126] After obtaining the attention matrix S, the spatial attention feature map P is calculated using the following formula:

[0127]

[0128] In the formula: Let be the feature value at the j-th position of feature map P; D is the feature map of size c×h×w obtained after the original feature map is convolved. Let the feature value at the j-th position of the input feature map A be denoted as ; Parameters used to control the weights of eigenvalues;

[0129] The channel attention matrix X is calculated by combining the input feature map with the attention mechanism. The X matrix expresses the attention magnitude of each channel to other channels, and its calculation formula is as follows:

[0130]

[0131] In the formula: Let the feature value be the feature value at the i-th position of the input feature map A; Let X be the eigenvalue of the j-th row and i-th column of matrix X;

[0132] After obtaining the attention matrix X, the channel attention feature map E can be calculated using the following formula:

[0133]

[0134] In the formula: Let J be the feature value at the j-th position of feature map E; Parameters used to control the weights of eigenvalues;

[0135] After obtaining the spatial attention feature map and the channel attention feature map, the two feature maps are added element by element to obtain a new set of feature maps with dimensions c×h×w.

[0136] Step 4: Use the UAV imagery slope disaster sample library to train the slope disaster information extraction model in real time on the deep learning workstation to obtain the trained model parameters.

[0137] First, training data is input, and the optimization algorithm continuously adjusts the model parameters to minimize the loss function on the training data. The trained model is then validated using a validation dataset to check its generalization ability and prediction accuracy. If the validation results are not ideal, the model's hyperparameters, such as the learning rate, regularization coefficient, and batch size, need to be adjusted to improve the model's performance and generalization ability. After training, the trained model parameters are recorded for deployment.

[0138] Step 5: As Figure 4As shown, the slope extraction model is quantized and compressed. Guided by minimizing relative entropy, the model's weight data type FP32 is nonlinearly mapped to INT8. This significantly reduces memory access and computation during model inference while maintaining model accuracy to the greatest extent, thereby accelerating model inference. The model quantization and compression method is as follows:

[0139] First, determine the pruning threshold for model quantization, remove outliers, and calculate the pruning threshold using the following formula:

[0140]

[0141] In the formula: These are the weight parameters for the FP32 type model; This is the clipping threshold.

[0142] The quantization scaling factor is calculated by using a pruning threshold. The floating-point dynamic range is divided into 256 intervals. The formula for calculating the scaling factor is:

[0143]

[0144] In the formula: is the scale factor.

[0145] Finally, guided by the principle of minimizing relative entropy, the model is quantized using a pruning threshold and a scaling factor, and the model's weight data type FP32 is nonlinearly mapped to INT8.

[0146] The formula is as follows:

[0147]

[0148] In the formula: These are model weight parameters of type INT8.

[0149] Due to the difference in the range of floating-point and integer values, the above quantization operations may result in a loss of accuracy. Therefore, calibration can be used to adjust quantization parameters to improve accuracy and performance. To obtain a stable INT8 quantization model, TensorRT's calibration function automatically calculates appropriate quantization parameters based on a pre-collected dataset of 1000 images, resulting in a reliable INT8 quantization model.

[0150] Step 6: The slope disaster information is extracted in real time, quantized, compressed, and deployed on an embedded device. On the platform, the camera module captures UAV images in real time, which are then transmitted to the embedded device for data preprocessing before being fed into the model for inference. The model completes processing before the camera module captures the next image, and all images are processed synchronously during flight. Finally, the complete semantic segmentation map, the original images captured by the camera module, and the location information are transmitted to the UAV ground receiver for visualization processing to intuitively observe the image information and disaster situation.

[0151] In one specific embodiment, the present invention also provides the following technical solution: a device, the device including at least one processor; and at least one processor is equipped with the aforementioned lightweight slope disaster sensing system.

[0152] The semantic segmentation process on this device is as follows: Each image is cropped into 512×512 blocks and then input into the model for inference. After the image is downsampled four times by the backbone network, it enters the global feature extraction and non-local feature extraction modules to obtain two new feature maps. These are then upsampled to 512×512 size and stitched together. Finally, a probability map is generated through convolution to obtain the semantic segmentation map. All the segmented semantic segmentation maps are stitched together to obtain the complete semantic segmentation map.

[0153] In one specific embodiment, the present invention also provides the following technical solution: a platform, such as... Figure 5 As shown, the platform includes the aforementioned device, camera module, positioning module, power supply module, and communication module;

[0154] Camera module: Used to acquire remote sensing images of the ground during the drone's flight;

[0155] Positioning module: Used to obtain the current position of the drone, add position information to the drone imagery, and improve the accuracy of the image data;

[0156] Communication module: Used to receive flight control signals and camera control signals, and transmit slope information extraction results and images captured by the camera module back to the UAV ground control terminal;

[0157] Power supply module: Used to supply power to the equipment, camera module, positioning module, and communication module.

[0158] The device is equipped with the aforementioned lightweight slope disaster sensing system.

[0159] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lightweight slope disaster perception system based on deep learning, characterized in that, The system includes: Slope disaster sample database construction module (110): Collect UAV image data of the disaster area, annotate the slope disaster information in it, and then form a slope disaster sample database after data preprocessing; Real-time slope disaster information extraction model construction module (120): used to construct a real-time slope disaster information extraction model; Global feature extraction module and non-local feature extraction module (130): used to construct the global feature extraction module and non-local feature extraction module to obtain more comprehensive contextual semantic information. The global feature extraction module and non-local feature extraction module are connected in parallel with the backbone network of the real-time slope disaster information extraction model as an extension of the backbone network. Real-time slope disaster information extraction model training module (140): The real-time slope disaster information extraction model is trained using a slope disaster sample library, and the model parameters are adjusted; Quantization and compression module (150): Quantizes and compresses the real-time slope disaster information extraction model, using the minimum relative entropy as a guide, nonlinearly mapping the model's weight data type FP32 to INT8, and quantizing and compressing the model by calculating the pruning threshold and scale factor; the quantization and compression includes the following steps: First, determine the pruning threshold for model quantization and remove outliers. The formula for calculating the pruning threshold is as follows: ； In the formula: These are the weight parameters for the FP32 type model; This is the clipping threshold; Then, the quantization scaling factor is calculated by using a pruning threshold. The floating-point dynamic range is divided into 256 intervals. The formula for calculating the scaling factor is as follows: ； In the formula: Scale factor; Finally, guided by the principle of minimizing relative entropy, the model is quantized using a pruning threshold and a scaling factor, and the model's weight data type FP32 is non-linearly mapped to INT8, as shown in the following formula: ； In the formula: These are model weight parameters of type INT8.

2. The lightweight slope disaster perception system based on deep learning according to claim 1, characterized in that: The backbone network structure of the real-time slope disaster information extraction model (120) has 101 layers, including 4 convolutional combinations, a global average pooling layer, a fully connected layer and a softmax layer. Each convolutional combination includes 3, 4, 23 and 3 residual modules in sequence. Each residual module contains 3 convolutional layers and 1 regularization layer.

3. The lightweight slope disaster perception system based on deep learning according to claim 1, characterized in that: The data preprocessing includes rotation, scaling, random cropping, and color transformation.

4. The lightweight slope disaster perception system based on deep learning according to claim 1, characterized in that: The global feature extraction module includes a global average pooling layer and two fully connected layers, comprising pooling operations, convolution operations, ReLU activation function, and sigmoid activation function; the specific calculations of this module include the following: First, global average pooling is used to aggregate the feature values ​​for each channel. This operation adds up all feature values ​​and calculates the average, as shown in the following formula: ； In the formula: H represents the statistical value obtained after global average pooling of the c-th channel of the feature map; H represents the height of the feature map; W represents the width of the feature map. This represents the c-th channel of the feature map; Secondly, the relationship between different channels is captured using two different fully connected layers. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the sigmoid activation function. The calculation formula is shown below: ； In the formula: s is the weight parameter obtained from learning different channels; Represents the sigmoid activation function; This represents the weight parameters of the first fully connected layer; This represents the weight parameters of the second fully connected layer; Finally, the activation values ​​of each channel learned by the global feature extraction module are multiplied by the original features at the corresponding positions on the input feature map, as shown in the following formula: ； In the formula: This represents the new feature map with dimensions c×h×w obtained after processing by the global feature extraction module.

5. The lightweight slope disaster perception system based on deep learning according to claim 1, characterized in that: The non-local feature extraction module includes a spatial attention module and a channel attention mechanism module, which consists of convolution operations, a softmax function, a reshape function, and a transpose function; the specific calculations of this module include the following: By combining the input feature map with an attention mechanism, the attention matrix S is calculated. This matrix expresses the weight of attention at each location to other locations on the feature map. The formula for calculating the attention matrix S is as follows: ； In the formula: B and C are the new feature maps of size c×h×w obtained after the original feature maps are convolved; The first of the S matrix j Line 1 i The characteristic values ​​of the column; After obtaining the attention matrix S, the spatial attention feature map P is calculated using the following formula: ； In the formula: Let be the feature value at the j-th position of feature map P; D is the feature map of size c×h×w obtained after the original feature map is convolved. Let the feature value at the j-th position of the input feature map A be denoted as ; Parameters used to control the weights of eigenvalues; The channel attention matrix X is calculated by combining the input feature map with the attention mechanism. The X matrix expresses the attention magnitude of each channel to other channels, and its calculation formula is as follows: ； In the formula: Let the feature value be the feature value at the i-th position of the input feature map A; Let X be the eigenvalue of the j-th row and i-th column of matrix X; After obtaining the attention matrix X, the channel attention feature map E is calculated using the following formula: ； In the formula: Let J be the feature value at the j-th position of feature map E; Parameters used to control the weights of eigenvalues; After obtaining the spatial attention feature map and the channel attention feature map, the two feature maps are added element by element to obtain a new set of feature maps with dimensions c×h×w.

6. A device, characterized in that: The device includes at least one processor; and at least one processor is equipped with the lightweight slope hazard sensing system according to any one of claims 1-5.

7. A platform, characterized in that: The platform includes the device, camera module, positioning module, power supply module, and communication module as described in claim 6; Camera module: Used to acquire remote sensing images of the ground during the drone's flight; Positioning module: Used to obtain the current position of the drone, add position information to the drone imagery, and improve the accuracy of the image data; Communication module: Used to receive flight control signals and camera control signals, and transmit slope information extraction results and images captured by the camera module back to the UAV ground control terminal; Power supply module: Used to supply power to the equipment, camera module, positioning module, and communication module.

Images