Landslide surface inspection method based on RTDETR

Abstract

The invention discloses a landslide surface inspection method based on RTDETR, and the method comprises the steps: collecting and obtaining an existing open-source landslide type data set, dividing the existing open-source landslide inspection data set into a training set, a verification set and a test set, and carrying out the specific scene data enhancement and conventional data enhancement operation of the data of the training set; constructing a landslide surface inspection model which is jointly formed by a ResNet-based multi-scale adaptive convolutional feature extraction backbone network, a CNN + Transform-based multi-scale feature fusion layer, and a DETR decoding layer and an adaptive high-order matching cost function-based detection head frame; learning weight parameters by using the training set, further optimizing and adjusting the verification set, and finally testing the landslide surface inspection model by using the test set to determine a final detection result; the accuracy of landslide type detection is greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of target detection, in particular to a landslide surface inspection method based on RTDETR. Background Art

[0002] Landslide geological disasters not only cause huge economic losses, but also lead to significant deformation of the surface. Landslide types can be mainly divided into landslide cracks, ground subsidence and uplift, and large stone rolling. Timely and accurate determination of landslide types is not only the basis for disaster analysis and judgment, but also the basis for decision-making in emergency rescue and engineering management.

[0003] Remote sensing technology has the advantages of being fast and flexible, having short data processing time, low technical cost and high spatial and temporal resolution of remote sensing images in the process of acquiring remote sensing data of landslide disasters. It is increasingly widely used in geological disaster investigations such as landslides, emergency rescue and other work, and has achieved good results.

[0004] Since the landslide inspection task has high real-time requirements for the detector, most of the existing landslide inspection work is carried out based on the traditional YOLO target detection algorithm. Although some detectors have achieved good results, there are several shortcomings: First, the YOLO series of algorithms are not end-to-end target detectors, and post-processing operations such as non-maximum suppression need to be introduced. The manual adjustment of the threshold greatly affects the performance of the detector; second, the structure based entirely on CNN will make the detector pay more attention to local details and relatively ignore global semantic information, especially in the process of multiple downsampling through the feature fusion module, a large amount of detail information will be lost. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a landslide surface inspection method based on RTDETR to detect the landslide surface type in real time and end-to-end, and obtain a more reliable detection result by efficiently combining the global and local semantic information of CNN and Transformer integrated features.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a landslide surface inspection method based on RTDETR, which comprises the following steps: Step S1, collect and obtain the existing open source landslide surface inspection data set, and divide it into training set, verification set and test set according to a certain ratio, and perform specific scene data enhancement and conventional data enhancement operations on the divided training set; Step S2, constructing a landslide surface inspection model composed of a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, a multi-scale feature fusion layer based on CNN+Transformer, and a detection head framework based on a DETR decoding layer and an adaptive high-order matching cost function; Step S3, sending the training set to the landslide surface inspection model obtained in step S2 for training, extracting features by a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, fusing multi-scale features by a multi-scale feature fusion layer based on CNN+Transformer, and then sending the features to a decoder and a detection head layer to predict the target category and position, finally matching the predicted result with the true label and calculating the cost, performing gradient descent on the training error with the cost value, completing the training weight parameter learning, and obtaining a trained landslide surface inspection model; Step S4, sending the verification set to the landslide surface inspection model trained in step S3, further estimating the generalization error, and adjusting the hyperparameters of the landslide surface inspection model; Step S5: Use the landslide surface inspection model test set after the hyperparameter adjustment optimization to complete the landslide surface inspection detection, evaluate the test results, and then use the landslide surface inspection model that has passed the test to detect the inspection data set.

[0007] Furthermore, in step S1, the specific scene data enhancement and conventional data enhancement operations are performed on the divided training set, including data enhancement of the images in the open source landslide surface inspection training set. The specific process is as follows: First, two types of noise are introduced, namely, elongated noise and arc or spherical noise, according to a certain ratio. Add two kinds of noise, is the ratio of the number of slender noise to the number of other noises; then specific noise is added to 45% of the images in the training set. If the image only has landslide cracks, Randomly add noise. If there is no landslide crack in the image, Add noise; perform traditional random photometric distortion, random cropping, random horizontal flipping around the Y axis at a fixed ratio and scale the enhanced image to 640×640.

[0008] Furthermore, the ResNet-based multi-scale adaptive convolutional feature extraction backbone network in step S2 includes three branches: the traditional convolution branch , dilated convolution branch , dynamic snake convolution branch , after the data enhancement in step S1, the image is sent to the three branches respectively to obtain the feature , , , and then normalize the three features and send them to three independent fully connected layers to get the corresponding weight values , , Finally, the weighted addition and fusion are performed according to the proportion of each feature to obtain the final feature .

[0009] Furthermore, the multi-scale feature fusion layer based on CNN+Transformer in step S2 consists of two parts: an intra-scale feature interaction module and an inter-scale feature interaction module; Among them, the intra-scale feature interaction module is divided into two parts: small-scale feature interaction and large-scale feature fusion according to the size of the feature map. Specifically, the feature map is obtained by selecting the multi-scale adaptive convolutional feature extraction backbone network based on ResNet described in step S2. of , , Layer Features , , Perform multi-scale feature fusion; Flattened into feature embedding vectors, reconstructed into two-dimensional features after self-attention for in-scale interaction ; At the same time, the features , A large-scale feature attention module LSFA is sent to it, whose input is a two-dimensional feature map and whose output is also a two-dimensional feature map. Specifically, the input feature is passed through a fully connected layer to obtain a learnable parameter , flatten the features greater than or equal to this parameter into an embedding vector and obtain new features through self-attention and scale interaction, and then compare the new features with those less than The features are reconstructed into a two-dimensional feature map and directly concatenated with the original input features. The multi-channel interaction information is fused through the convolution layer to obtain the output two-dimensional feature map. , ; The feature interaction between scales is constructed based on PANet, and its input is the features after the interaction within the scale. , , , the output is the feature after the interaction between scales , , .

[0010] Furthermore, in step S2, the detection head based on the DETR decoding layer and the adaptive high-order matching cost function is used to generate a set Q containing the target; First, the three-layer features obtained after feature fusion , , Flattened into embedding vectors and concatenated together to form feature vectors , and send it to the detection head to obtain the location information of the corresponding query and its classification confidence score; then select the top 300 queries with the above confidence scores as the initial target , at each decoder layer Perform self-attention operation and then interact with the feature vector between scales Do cross-attention to get a new set of prediction targets , helping the model learn the location and feature representation of the target object in the image; Then the target set corrected by the decoder is Perform binary matching with Groundtruth and use the Hungarian algorithm for one-to-one priority matching to find the category and location information corresponding to the minimum matching cost. When the matching cost function is a linear combination of the classification score and the positioning score, the number of iterations is greater than or equal to The matching cost function uses an adaptive high-order matching cost function ,in, is the classification confidence, Represents the intersection-over-union ratio of the predicted box and the true box, and is a hyperparameter, which is set differently in different iteration cycles. and , prompting the model to pay more attention to local semantic information in the early stage and more attention to global semantic information in the later stage.

[0011] Furthermore, in step S3, the training error is gradient-decreased by the cost value to complete the specific process of learning the trainable weight parameters as follows: By optimizing the parameters of the model through the cost function and gradient descent method, the landslide surface inspection model can gradually reduce the training error and improve the prediction accuracy. By using the Adam optimizer and learning rate adjustment, the model can converge efficiently during the training process and finally obtain the landslide surface inspection model.

[0012] Furthermore, the specific process of estimating the generalization error and adjusting the hyperparameters of the landslide surface inspection model in step S4 is as follows: The generalization ability of the model is verified using multiple indicators such as classification accuracy, average precision and recall rate to avoid overfitting of training data and ensure the stability and accuracy of the model on new data. At the same time, multiple hyperparameters such as learning rate, batch size and network structure are adjusted to ensure model performance while improving its adaptability to landslide surface inspection tasks.

[0013] Furthermore, the specific process of the landslide surface inspection model that has passed the test in step S5 is as follows: The model is comprehensively evaluated on the test set to confirm its accuracy and generalization ability, ensuring that it can effectively identify the type of landslide surface in practical applications. Finally, the qualified model is deployed on the actual data set for landslide inspection tasks to obtain usable and accurate detection results.

[0014] Beneficial effects of the present invention: 1. Aiming at the problem that the YOLO detection algorithm commonly used in industry is not end-to-end, the present invention constructs a target detection method for landslide surface inspection based on the deep learning method of the DETR general detection paradigm; this method improves the performance of the basic detection framework from multiple angles and improves the accuracy of landslide type detection.

[0015] 2. The present invention adopts a multi-scale adaptive convolutional feature extraction network based on ResNet, and uses different convolution types to extract features in a targeted manner, which increases the "width" of the backbone network to a certain extent, enhances the expression ability and nonlinear mapping ability of the model, and reduces the problem of gradient disappearance.

[0016] 3. In order to reduce the differences in the process of feature interaction between scales, the present invention proposes a large-scale feature attention module (LSFA), which enables feature maps with larger scales to perform feature interaction within the scale through the attention mechanism, while enhancing the correlation between feature interactions between scales.

[0017] 4. In the final binary matching process, the present invention proposes a phased matching cost function to guide the model to focus on different semantic information at different times, and gradually shift the learning of local semantic information to global detail semantic information to enhance the learning and representation capabilities of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a landslide surface inspection method based on RTDETR according to an embodiment of the present invention; Figure 2 A schematic diagram of a ResNet multi-scale adaptive convolutional feature extraction network according to an embodiment of the present invention; Figure 3 A schematic diagram of the overall framework of the detection model and the constructed large-scale feature attention module LSFA according to an embodiment of the present invention; Figure 4 It is a visual display result of testing the technical solution of the present invention and the comparative model. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0020] Example 1: Reference Figure 1The embodiment of the present invention provides a landslide surface inspection method based on RTDETR, comprising the following steps: Step S1, collect and obtain the existing open source landslide surface inspection (landslide cracks, large rock rolling) data set, and divide it into training set, verification set and test set according to a certain ratio, and perform specific scene data enhancement and general data enhancement operations on the divided training set; The open source landslide surface inspection dataset is divided into training set, validation set and test set (ratio is 6:2:2). The closer the image data of each category is, the better the training effect is. Data enhancement is performed on the images in the training set, including: firstly, two types of noise (slender noise, arc or spherical noise) are introduced, and then the noise is divided into two types of noise (slender noise, arc or spherical noise) according to a certain ratio. Add two kinds of noise ( is the ratio of the number of slender noise to the number of other noise); then add specific noise to 45% of the images in the training set. If the image only has landslide cracks, Randomly add noise, if there is no landslide crack in the image, Add noise; perform traditional random photometric distortion, random cropping, random horizontal flipping around the Y axis at a fixed ratio and scale the enhanced image to 640×640.

[0021] Step S2: construct a landslide surface inspection model consisting of a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, a multi-scale feature fusion layer based on CNN+Transformer, and a detection head framework based on a DETR decoding layer and an adaptive high-order matching cost function (refer to Figure 3 ); The ResNet-based multi-scale adaptive convolutional feature extraction backbone network module (refer to Figure 2 ) consists of three branches: traditional convolution branch , dilated convolution branch , dynamic snake convolution branch , after the data enhancement in step S1, the image is sent to the three branches respectively to obtain the feature , , , and then normalize the three features and send them to three independent fully connected layers to get the corresponding weight values , , Finally, the weighted addition and fusion are performed according to the proportion of each feature to obtain the final feature .

[0022] The multi-scale feature fusion layer based on CNN+Transformer (refer to Figure 3) consists of two parts: intra-scale feature interaction module and inter-scale feature interaction module; Among them, the intra-scale feature interaction is divided into two parts according to the size of the feature map: small-scale feature interaction and large-scale feature fusion. Specifically, the feature extraction backbone network based on the ResNet multi-scale adaptive convolution is selected. of , , Layer Features , , Perform multi-scale feature fusion. Flattened into feature embedding vectors, reconstructed into two-dimensional features after self-attention for in-scale interaction ; At the same time, the features , A large-scale feature attention module (LSFA) is sent to it, whose input is a two-dimensional feature map and whose output is also a two-dimensional feature map. Specifically, the input feature is passed through a fully connected layer to obtain a learnable parameter , flatten the features greater than or equal to this parameter into an embedding vector and obtain new features through self-attention and scale interaction, and then compare the new features with those less than The features are reconstructed into a two-dimensional feature map and directly concatenated with the original input features. The multi-channel interaction information is fused through the convolution layer to obtain the output two-dimensional feature map. , ; The feature interaction between scales is constructed based on PANet, and its input is the features after the interaction within the scale. , , , the output is the feature after the interaction between scales , , ; Specifically, the input features After upsampling with 1×1 convolution, SiLU activation function and bilinear interpolation, the feature Splice them together and then use a Fusion module to get the fused features , and then repeat this feature through 1×1 convolution, SiLU activation function and bilinear interpolation upsampling and feature Spliced ​​together and fused through the Fusion module to obtain features ; Then the feature After 3×3 convolution and SiLU activation function, the feature The upsampled features are concatenated together and fused through the Fusion module to obtain the features. , and similarly the features After 3×3 convolution and SiLU activation function, the feature The upsampled images are spliced ​​together and the features are obtained by fusion module. Among them, the Fusion module is mainly composed of two branches, one of which is only a 1×1 convolution layer, and the other is composed of a 1×1 convolution and multiple convolution blocks.

[0023] The DETR-based decoding layer first combines the three layers of features obtained by feature fusion , , Flattened into embedding vectors and concatenated together to form feature vectors , and send it to the detection head to obtain the location information of the corresponding query and its classification confidence score. Then select the top 300 queries with the above confidence scores as the initial target , at each decoder layer Perform self-attention operation and then interact with the feature vector between scales Do cross-attention to get a new set of prediction targets , helping the model learn the location and feature representation of the target object in the image.

[0024] The detection head of the adaptive high-order matching cost function specifically converts the target set corrected by the decoder into Perform binary matching with Groundtruth and use the Hungarian algorithm for one-to-one priority matching to find the category and location information corresponding to the minimum matching cost. When the matching cost function is a linear combination of the classification score and the positioning score, the number of iterations is greater than or equal to The matching cost function uses an adaptive high-order matching cost function ,in, is the classification confidence, Represents the intersection-over-union ratio of the predicted box and the true box, and It is a hyperparameter, which is set differently in different iteration cycles. and , prompting the model to pay more attention to local semantic information in the early stage and more attention to global semantic information in the later stage.

[0025] Find the category and location information corresponding to the minimum matching cost and output them to obtain the corresponding detection result; finally, input the test set into the model, and use indicators such as FPS and mAP to evaluate the detection effect of the model.

[0026] Step S3, sending the training set to the landslide surface inspection model obtained in step S2 for training, extracting features by a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, fusing multi-scale features by a multi-scale feature fusion layer based on CNN+Transformer, and then sending the features to a detection head based on a DETR decoding layer and an adaptive high-order matching cost function to predict the target category and position, finally matching the predicted result with the true label and calculating the cost, performing gradient descent on the training error with the cost value, completing the learning of trainable weight parameters, and obtaining a trained landslide surface inspection model.

[0027] Step S4: sending the verification set to the landslide surface inspection model trained in step S3 to further estimate the generalization error and adjust the hyperparameters of the landslide surface inspection model.

[0028] Step S5, using the landslide surface inspection model test set after the hyperparameter adjustment optimization, complete the detection of the landslide surface inspection, evaluate the test results, and then use the qualified landslide surface inspection model to detect the landslide surface inspection data set.

[0029] In the present invention, the ResNet-based multi-scale adaptive convolutional feature extraction backbone network module consists of three branches: the traditional convolution branch , dilated convolution branch , dynamic snake convolution branch The present invention adopts dynamic snake-shaped convolution to capture the features of slender landslide cracks, uses hole convolution to expand the receptive field to capture the features of large stone rolling that require more global semantic information, and uses ordinary convolution to extract basic features. Finally, the three normalized features are weightedly added and fused according to the proportion of each feature to complement the information between the feature semantic information.

[0030] In the present invention, the multi-scale feature fusion layer based on CNN+Transformer includes an intra-scale feature interaction module. For feature maps with smaller scales, the intra-scale information interaction is performed directly through the Transformer; for feature maps with larger scales, a learnable parameter is obtained through adaptive fully connected layers. , the features greater than this parameter are regarded as foregrounds with rich semantic information, and the features less than this parameter are regarded as backgrounds without semantic information, so as to maintain the consistency between feature values ​​for subsequent inter-scale feature interactions. The Transformer is used to perform intra-scale feature interactions on the foreground with rich semantic information, and then the two-dimensional features reconstructed from them and the background feature elements are fused with multi-channel interaction information through the convolutional layer.

[0031] In the present invention, the multi-scale feature fusion layer based on CNN+Transformer includes an inter-scale feature interaction module, which is a fusion method constructed based on the path aggregation network PANet, and can effectively enhance the semantic classification ability of all features.

[0032] In the present invention, when initializing the target set Q, the DETR-based decoding layer selects a set consisting of the top 300 queries in classification confidence scores, and then performs cross-attention with the features after interaction with the inter-scale features to correct the target set position and category information.

[0033] In the present invention, the detection head of the adaptive high-order matching cost function still adopts the binary matching, but the difference is that the present invention adopts different cost functions in different training cycles; when the number of training iterations is less than When , the cost function is constructed according to the linear combination of the classification score and the positioning score, and the number of iterations is greater than or equal to The matching cost function uses an adaptive high-order matching cost function ,in, is the classification confidence, Represents the intersection-over-union ratio of the predicted box and the true box, and is a hyperparameter, we can set is 1, The value gradually increases with the number of iterations, because the classification confidence and The values ​​are between 0 and 1, so as Value increases, The value of is relatively reduced, and the model's attention is shifted from local semantic information to global information, further enhancing the model's learning ability.

[0034] The step S3 gradient descent learns the trainable weight parameters, and the specific process is as follows: the parameters of the model are optimized by the cost function and the gradient descent method, so that the landslide surface inspection model can gradually reduce the training error and improve the prediction accuracy. By using advanced optimization algorithms (such as Adam optimizer) and appropriate learning rate adjustment strategies, the model can converge efficiently during the training process, and finally a landslide surface inspection model with high accuracy and strong robustness is obtained; The step S4 estimates the generalization error and adjusts the hyperparameters, and the specific process is as follows: using classification accuracy, mean average precision (mAP) and recall rate to verify the generalization ability of the model, avoid overfitting the training data, ensure the stability and accuracy of the model on new data, and adjust multiple hyperparameters such as learning rate, batch size and network structure to ensure the performance of the model while improving its adaptability to the landslide surface inspection task; The step S5 tests the trained model, and the specific process is as follows: comprehensively evaluate the model on the test set to confirm the accuracy and generalization ability of the model to ensure that it can effectively identify the type of landslide surface in practical applications, and finally deploy the qualified model to the actual data set for landslide inspection tasks to obtain usable and accurate detection results.

[0035] To verify the effectiveness of the present invention, we use a test set to test the technical solution and comparison model of the present invention, and visualize some of the results in Figures (a), (b), and (c), where Figure (a) shows the detection results of YOLOV8 on two groups of landslide surface cracks and large rockfall images, and Figure (b) shows the detection results of the original RTDETR on two groups of landslide surface cracks and large rockfall images. The RTDETR* in Figure (c) represents the improved RTDETR proposed in the present invention, which also shows its detection results for two groups of landslide surface cracks and large rockfall images. It can be seen that the improved RTDETR is more accurate in target classification and positioning in complex environments, and can effectively detect the landslide surface, verifying the effectiveness of the present invention.

[0036] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A landslide surface inspection method based on RTDETR, characterized in that: It includes the following steps: Step S1, collect and obtain the existing open source landslide surface inspection data set, and divide it into training set, verification set and test set according to a certain ratio, and perform specific scene data enhancement and conventional data enhancement operations on the divided training set; Step S2, constructing a landslide surface inspection model composed of a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, a multi-scale feature fusion layer based on CNN+Transformer, and a detection head framework based on a DETR decoding layer and an adaptive high-order matching cost function; Step S3, sending the training set to the landslide surface inspection model obtained in step S2 for training, extracting features by a multi-scale adaptive convolutional feature extraction backbone network based on ResNet, fusing multi-scale features by a multi-scale feature fusion layer based on CNN+Transformer, and then sending the features to a decoder and a detection head layer to predict the target category and position, finally matching the predicted result with the true label and calculating the cost, performing gradient descent on the training error with the cost value, completing the training weight parameter learning, and obtaining a trained landslide surface inspection model; Step S4, sending the verification set to the landslide surface inspection model trained in step S3, further estimating the generalization error, and adjusting the hyperparameters of the landslide surface inspection model; Step S5: Use the landslide surface inspection model test set after the hyperparameter adjustment optimization to complete the landslide surface inspection detection, evaluate the test results, and then use the landslide surface inspection model that has passed the test to detect the inspection data set.

2. The landslide surface inspection method based on RTDETR according to claim 1 is characterized in that: In step S1, the divided training set is subjected to scene-specific data enhancement and conventional data enhancement operations, including data enhancement of images in the open-source landslide surface inspection training set. The specific process is as follows: First, two types of noise are introduced, namely, elongated noise and arc or spherical noise, according to a certain ratio. Add two kinds of noise, is the ratio of the number of slender noise to the number of other noises; then specific noise is added to 45% of the images in the training set. If the image only has landslide cracks, Randomly add noise. If there is no landslide crack in the image, Add noise; perform traditional random photometric distortion, random cropping, random horizontal flipping around the Y axis at a fixed ratio and scale the enhanced image to 640×640.

3. A landslide surface inspection method based on RTDETR according to claim 2, characterized in that: The ResNet-based multi-scale adaptive convolution feature extraction backbone network in step S2 includes three branches: the traditional convolution branch , dilated convolution branch , dynamic snake convolution branch , after the data enhancement in step S1, the image is sent to the three branches respectively to obtain the feature , , , and then normalize the three features and send them to three independent fully connected layers to get the corresponding weight values , , Finally, the weighted addition and fusion are performed according to the proportion of each feature to obtain the final feature .

4. The landslide surface inspection method based on RTDETR according to claim 3 is characterized in that: The multi-scale feature fusion layer based on CNN+Transformer in step S2 consists of two parts: an intra-scale feature interaction module and an inter-scale feature interaction module; Among them, the intra-scale feature interaction module is divided into two parts according to the size of the feature map: small-scale feature interaction and large-scale feature fusion. Specifically, the feature map is obtained by selecting the multi-scale adaptive convolutional feature extraction backbone network based on ResNet described in step S2. of , , Layer Features , , Perform multi-scale feature fusion; Flattened into feature embedding vectors, reconstructed into two-dimensional features after self-attention for in-scale interaction ; At the same time, the features , A large-scale feature attention module LSFA is sent to it, whose input is a two-dimensional feature map and whose output is also a two-dimensional feature map. Specifically, the input feature is passed through a fully connected layer to obtain a learnable parameter , flatten the features greater than or equal to this parameter into an embedding vector and obtain new features through self-attention and scale interaction, and then compare the new features with those less than The features are reconstructed into a two-dimensional feature map and directly concatenated with the original input features. The multi-channel interaction information is fused through the convolution layer to obtain the output two-dimensional feature map. , ; The feature interaction between scales is constructed based on PANet, and its input is the features after the interaction within the scale. , , , the output is the feature after the interaction between scales , , .

5. The landslide surface inspection method based on RTDETR according to claim 4 is characterized in that: In step S2, the detection head based on the DETR decoding layer and the adaptive high-order matching cost function is used to generate a set Q containing the target; First, the three-layer features obtained after feature fusion , , Flattened into embedding vectors and concatenated together to form feature vectors , and send it to the detection head to obtain the location information of the corresponding query and its classification confidence score; then select the top 300 queries with the above confidence scores as the initial target , at each decoder layer Perform self-attention operation and then interact with the feature vector between scales Do cross-attention to get a new set of prediction targets , helping the model learn the location and feature representation of the target object in the image; Then the target set corrected by the decoder is Perform binary matching with Groundtruth and use the Hungarian algorithm for one-to-one priority matching to find the category and location information corresponding to the minimum matching cost; during the training process, when the number of iterations is less than When the matching cost function is a linear combination of the classification score and the positioning score, the number of iterations is greater than or equal to The matching cost function uses an adaptive high-order matching cost function ,in, is the classification confidence, Represents the intersection-over-union ratio of the predicted box and the true box, and is a hyperparameter, which is set differently in different iteration cycles. and , prompting the model to pay more attention to local semantic information in the early stage and more attention to global semantic information in the later stage.

6. The landslide surface inspection method based on RTDETR according to claim 1 is characterized in that: In step S3, the specific process of performing gradient descent on the training error with the cost value to complete the learning of the trainable weight parameters is as follows: By optimizing the parameters of the model through the cost function and gradient descent method, the landslide surface inspection model can gradually reduce the training error and improve the prediction accuracy. By using the Adam optimizer and learning rate adjustment, the model can converge efficiently during the training process and finally obtain the landslide surface inspection model.

7. The landslide surface inspection method based on RTDETR according to claim 1 is characterized in that: The specific process of estimating the generalization error and adjusting the hyperparameters of the landslide surface inspection model in step S4 is as follows: The generalization ability of the model is verified using multiple indicators such as classification accuracy, average precision and recall rate to avoid overfitting of training data and ensure the stability and accuracy of the model on new data. At the same time, multiple hyperparameters such as learning rate, batch size and network structure are adjusted to ensure model performance while improving its adaptability to landslide surface inspection tasks.

8. The landslide surface inspection method based on RTDETR according to claim 1 is characterized in that: The specific process of testing the qualified landslide surface inspection model in step S5 is as follows: The model is comprehensively evaluated on the test set to confirm its accuracy and generalization ability, ensuring that it can effectively identify the type of landslide surface in practical applications. Finally, the qualified model is deployed on the actual data set for landslide inspection tasks to obtain usable and accurate detection results.

Images