A dam apparent disease detection method based on improved DSD-DETR

Abstract

The application discloses a dam apparent disease detection method based on an improved DSD-DETR, relates to the field of hydraulic engineering health monitoring, takes an RT-DETR model as a benchmark to design an EM-BasicBlock residual module, fuses deep convolution and star operation, and improves the upper limit of dam surface disease feature expression; reconstructs an encoder into an MSFF module, expands two parallel double-path row structures through point-by-point convolution operation, introduces learnable weight and bias variables, and better balances local details and global structure in a complex background; and modifies an original loss function into a dynamic balance loss function, improving regression optimization efficiency and stability in the training stage. Through the above improvement, the application realizes the lightweight of the model while enhancing the detection efficiency in the complex background, balances the detection accuracy and real-time performance, and provides an efficient solution for concrete dam surface disease target detection in a complex background.

Description

Technical Field

[0001] This invention relates to the field of health monitoring of water conservancy projects, and in particular to a method for detecting apparent defects in dams based on an improved DSD-DETR. Background Technology

[0002] Concrete dams, as key facilities in water conservancy projects, undertake multiple functions such as water resource regulation, flood control, irrigation, and power generation. Their safety and stability are directly related to the socio-economic development of the surrounding areas and the safety of people's lives and property. However, with the increase in service time and changes in the external environment, concrete dams gradually expose various surface defects, mainly including cracks, spalling, seepage, exudates, and spalling. The long-term existence of these defects poses a threat to the safety of the dam, and effective detection and timely repair measures must be taken to ensure its long-term stable operation. At present, the detection of surface damage to dams still mainly relies on manual visual inspection, on-site inspection, and conventional video recording. Although these methods have the advantages of direct implementation and mature operation experience, they generally suffer from problems such as high labor intensity, strong subjectivity, low detection efficiency, poor consistency of results, and high risks of operation in dangerous areas such as high slopes, high dam surfaces, and humid areas in actual engineering applications.

[0003] In recent years, with the rapid development of deep learning and computer vision technologies, image recognition-based methods for detecting structural surface defects have received widespread attention. Advanced algorithms, such as the YOLO series, Mask R-CNN, and DETR-like models, can quickly locate and identify defects like cracks, spalling, and leakage by automatically extracting image features and performing multi-scale modeling. Compared to traditional manual inspection methods, these methods offer significant advantages in detection speed, automation, and batch processing capabilities, providing a new technological path for intelligent detection of surface defects in concrete dams.

[0004] However, most existing intelligent detection methods for dam surface defects are trained and validated under conditions of uniform lighting, relatively simple backgrounds, and high image quality, which still have significant limitations in real-world, complex engineering scenarios. The service environment of concrete dams is complex, and images acquired on-site are often affected by various factors such as uneven lighting, shadow occlusion, surface contamination, deposits, texture interference, and variations in shooting distance. These factors easily lead to blurred boundaries, lost details, and weakened local features in defect areas. RT-DETR does not perform well under such conditions. Furthermore, RT-DETR generally suffers from large parameter sizes, high computational costs, and insufficient ability to identify small targets and defects with weak features, making it difficult to balance detection accuracy, real-time performance, and engineering deployment requirements. Summary of the Invention

[0005] To address the problems of blurred boundaries, lack of detail, and weakened local features in existing dam surface defect detection methods under complex backgrounds, this invention proposes an improved DSD-DETR-based method for dam surface defect detection under complex backgrounds. The aim of this invention is to improve the accuracy of defect target detection and reduce model weight in complex backgrounds.

[0006] To achieve the above objectives, this invention provides a method for detecting apparent defects in dams based on an improved DSD-DETR, comprising the following steps: S1. Obtain the dataset of complex apparent defects of the dam, and preprocess the dataset, dividing the preprocessed dataset into training set, validation set and test set. S2. Using the RT-DETR model as the baseline model, replace the residual module of the baseline model with the EM-BasicBlock residual module, and improve the residual module of the RT-DETR model backbone network. S3. Replace the AIFI module of the baseline model with the MSFF module and improve the AIFI module in the RT-DETR model encoder; S4. Replace the loss function of the baseline model with a dynamic equilibrium loss function to improve the loss function of the RT-DETR model; S5. Train the improved DSD-DETR model using the training and validation sets; S6. Input the test set into the trained DSD-DETR model to obtain the detection results.

[0007] Preferably, the specific steps of the preprocessing in S1 are as follows: S11. Scale the original image to 640×640 pixels; S12. Perform image enhancement on the scaled original image. Image enhancement includes rotation, flipping, Gaussian blur, contrast adjustment, and brightness adjustment. S13. Label the disease information, and label cracks, peeling, water leakage, exudate and spalling as CR, SP, WL, FBD and EA respectively; S14. After the annotation is completed, divide the dataset into training set, validation set and test set in a ratio of 8:1:1.

[0008] Preferably, the specific steps of S2 are as follows: S21. Two identical fully connected layers are generated by depthwise convolution with a 7×7 kernel. S22. Two fully connected layers are combined into a completely new fully connected layer through a star operation; S23. Reduce computational load by using 1×1 convolutional compression channels to achieve feature extraction.

[0009] Preferably, the specific formula for the star diagram operation in S22 is as follows: ; in, d is the input channel number, which is expanded to accommodate multiple output channels. Simultaneously processing with Multiple feature elements.

[0010] Preferably, the EM-BasicBlock residual module in S2 only replaces the 6th and 7th layers of the RT-DETR model backbone network, and its optimal embedding depth and number are determined by comparing the detection accuracy, recall rate and mAP index under different layer configurations.

[0011] Preferably, the specific steps of S3 are as follows: S31. Input a feature tensor The channel dimension is increased through pointwise convolution operations; S32. Orient the unfolded tensor to two parallel branches and enhance the extraction of multi-scale damage information through extended 3×3 and normal 3×3 depth convolutions. S33. Feature enhancement is performed on both branches using learnable weights and bias variables; S34. After passing the inverse fast Fourier transform, the activation output on the branch of the extended 3×3 depth convolution is multiplied pixel by pixel with the direct output of the other branch. S35. Perform a 1×1 convolution on the fused features.

[0012] Preferably, the dynamic equilibrium loss function in S4 is the joint loss function Focaler-IoU and CIoU, Focaler-CIoU, and the specific formula of Focaler-CIoU is as follows: ; in, Focaler-CIoU is the joint loss function used for training optimization of bounding box regression. For complete intersection-over-union, it is used to comprehensively measure the degree of overlap between the predicted box and the ground truth box, the distance between the center points, and the consistency of the aspect ratio; Intersection over Union (IoU) is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box. This is the Focaler-IoU term, used to dynamically reweight the regression gradients of samples of different quality.

[0013] Therefore, the present invention employs the above-mentioned method for detecting apparent defects in dams based on improved DSD-DETR, which has the following beneficial effects: (1) The improved EM-BasicBlock residual module, MSFF module and dynamic balance loss function of this invention can effectively enhance the network's detection accuracy for complex disease targets. EM-BasicBlock improves the upper limit of feature representation, MSFF improves the quality of cross-scale information fusion, and dynamic balance loss improves the efficiency and stability of regression optimization during the training phase. The three form a closed-loop gain from "feature extraction - feature fusion - target optimization", enabling the model to obtain stronger discrimination ability, more stable localization ability and better generalization performance under complex working conditions, thereby improving the overall detection accuracy and engineering applicability.

[0014] (2) The improved RT-DETR model of this invention has stronger end-to-end global modeling capabilities and lower hyperparameter dependence compared with the traditional single-stage detection framework. Compared with the benchmark model RT-DETR, this invention simultaneously optimizes the backbone feature representation, cross-scale fusion and regression objective function.

[0015] (3) This invention makes a series of improvements based on the RT-DETR model, taking advantage of its advantages in real-time performance and high accuracy to improve its robustness in the face of complex dam surface backgrounds and improve the identification accuracy of dam surface damage detection. Compared with traditional manual detection of concrete dam surface damage, this invention can significantly improve the accuracy of concrete dam surface damage detection and save a lot of time, material resources and financial resources.

[0016] (4) The detection method proposed in this invention achieves better results in terms of accuracy, recall and mean precision while reducing the number of parameters compared with the original RT-DETR-r18 model.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a network structure diagram of the DSD-DETR model in an embodiment of the present invention, which is based on an improved DSD-DETR method for detecting apparent defects in dams. Figure 2 The following is a structural diagram of the EM-BasicBlock residual module of an embodiment of the present invention, which is based on an improved DSD-DETR method for detecting dam apparent defects. (a) is the EM residual structure diagram, and (b) is the specific structural parameters. Figure 3 This is a structural diagram of the MSFF module in an embodiment of the present invention, which is a method for detecting apparent defects in dams based on an improved DSD-DETR. Figure 4The graphs shown are curves during the training process of an embodiment of the dam apparent defect detection method based on improved DSD-DETR of the present invention. (a) is the index curve, (b) is the precision-recall curve, and (c) is the F1-confidence curve. Figure 5 This is a comparative heatmap of an embodiment of the present invention, which includes an improved module for detecting apparent defects in dams based on the improved DSD-DETR. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] Example Please see Figures 1-5 This invention provides a method for detecting apparent defects in dams based on an improved DSD-DETR, such as... Figure 1 As shown, a series of improvements were made using the RT-DETR model as a benchmark, leveraging its advantages in real-time performance and high accuracy to improve its robustness in complex dam surface environments. The specific steps include: S1. Obtain the dataset of complex apparent defects of the dam and perform preprocessing.

[0022] The complex surface defects dataset includes images of cracks, spalling, seepage, exudates, and spalling. All images were captured using drones on the surface of the dam's spillway, resulting in multiple images of the dam surface. The acquired images underwent preprocessing, including scaling the original images to 640×640 pixels and image enhancement (rotation, flipping, Gaussian blur, contrast and brightness adjustment, etc.). Defect information was manually labeled, assigning cracks, spalling, seepage, exudates, and spalling the labels CR, SP, WL, FBD, and EA, respectively. After labeling, the dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0023] The dataset used in this embodiment consists of images taken by drones on the spillway of a concrete dam. This data contains information on various types of damage, including cracks, spalling, seepage, exudates, and spalling. Preprocessing included image enhancement of the dataset. This embodiment used flipping, Gaussian blur, brightness enhancement, and contrast enhancement, with a Gaussian blur radius of 2.5, a brightness enhancement factor of 1.5, and a contrast factor of 1.2. The image size transmitted to the model was uniformly set to 640×640. Simultaneously, the experimental hyperparameters were set as follows: 200 training epochs, a batch size of 16, SGD as the optimizer, an initial learning rate of 0.01, an optimizer momentum of 0.734, and an optimizer weight decay coefficient of 0.00015.

[0024] S2. Residual module of the backbone network of the improved RT-DETR model.

[0025] The residual module of the baseline model (RT-DETR model) is replaced with a newly designed EM-BasicBlock residual module, such as... Figure 2 The EM-BasicBlock residual module shown first uses a 7×7 depthwise convolution to acquire the current feature information. Then, it generates two identical fully connected layers through convolution. These two fully connected layers are then combined into a new fully connected layer through a star-shaped operation. Finally, it compresses the channels through 1×1 convolutions to reduce computation, achieving efficient and powerful feature extraction. This design maintains the stable training of the original ResNet residual connections while possessing stronger regularization and generalization capabilities.

[0026] Star operations demonstrate their ability to fuse high-dimensional features. After stacking multiple layers of residual blocks containing star operations, the latent dimension can be significantly increased to extremely high levels. In a single-layer neural network, the star operation is... This indicates that the features of two linear transformations are combined by element-wise multiplication. Specifically, the star diagram operation is shown in the following formula: ; in, d is the input channel number, which can be easily expanded to accommodate multiple output channels. Simultaneously processing with Multiple feature elements.

[0027] The EM-BasicBlock residual module maps input features to a higher-dimensional nonlinear representation space through an enhanced mapping mechanism. Under the residual learning framework, it expands the expression range of feature functions, enabling shallow texture details and deep semantic information to be more fully preserved and transmitted. This improves the separability of disease targets with weak contrast, strong texture interference, slender shape, or blurred boundaries (such as cracks, rust edges, local erosion, etc.), thereby reducing false detections and false negatives under complex background conditions.

[0028] S3. Improve the AIFI module in the RT-DETR model encoder.

[0029] By replacing the AIFI module of the baseline model with the designed MSFF module, this invention introduces a dual-path, multi-range depth convolutional structure in the encoder to better balance local details and global structure. Simultaneously, drawing on the equivalence of convolution in the spatial and spectral domains, this invention proposes a multi-scale feature fusion module (MSFF). By introducing an adaptive bias filter in the spectral space, the feature frequency components are weighted and adjusted, effectively enhancing texture features related to dam damage and suppressing background redundancy. Therefore, the MSFF module enables more efficient separation and identification of dam surface damage, providing solid support for accurate detection in complex environments.

[0030] The MSFF module structure is as follows: Figure 3 As shown, input a feature tensor The channel dimension is increased through pointwise convolution operations. The unfolded tensor is then directed to two parallel branches, where multi-scale damage information extraction is enhanced through extended 3×3 and normal 3×3 depthwise convolutions. Feature enhancement is then performed on both branches using learnable weights and bias variables. After passing through an Inverse Fast Fourier Transform, the activation output of the branch using extended 3×3 depthwise convolution is multiplied pixel-by-pixel with the direct output of the other branch. Finally, the fused features are convolved using a 1×1 convolution.

[0031] The MSFF module addresses the typical characteristics of disease scale variations and irregular morphology. It enhances the encoder's effective representation capability through the complementary fusion of multi-scale information. On the one hand, it strengthens global information modeling to improve the overall perception of large-scale disease areas. On the other hand, it preserves key structural details as much as possible during downsampling, thereby simultaneously meeting the detection needs of both small and large targets and improving recall and localization stability in multi-scale scenarios.

[0032] S4. Improve the loss function of the RT-DETR model.

[0033] The loss function of the baseline model is replaced with a dynamic equilibrium loss function, which is a joint loss function of Focaler-IoU and CIoU. By incorporating Focaler-IoU into CIoU, a new bounding box loss function, called Focaler-CIoU, is obtained. The specific formula for Focaler-CIoU is: ; in, Focaler-CIoU is the joint loss function used for training optimization of bounding box regression. For complete intersection-over-union, it is used to comprehensively measure the degree of overlap between the predicted box and the ground truth box, the distance between the center points, and the consistency of the aspect ratio; Intersection over Union (IoU) is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box. This is the Focaler-IoU term, used to dynamically reweight the regression gradients of samples of different quality.

[0034] The dynamic balance loss function adopts the Focaler-CIoU joint loss to simultaneously constrain the overlap relationship and geometric consistency. It also dynamically adjusts the gradient contribution of samples of different quality through the "focusing" mechanism, which alleviates the optimization imbalance caused by easy samples dominating the model and strengthens the learning of difficult samples and low-quality matching samples. This makes the bounding box regression converge more stably and the localization accuracy higher, especially under complex backgrounds, occlusion interference, or unclear target boundaries.

[0035] S5. Use the training and validation sets prepared in S1 to train the DSD-DETR model improved by S2, S3 and S4.

[0036] In this invention, the EM-BasicBlock residual module replaces only layers 6 and 7 of the backbone network. By comparing the detection accuracy, recall, and mAP under different layer configurations, the role of the EM-BasicBlock residual block in multi-scale feature fusion can be revealed, thereby determining its optimal embedding depth and number. This experiment not only provides a theoretical basis for network structure design but also lays a practical foundation for improving the accuracy and robustness of dam surface defect detection.

[0037] As shown in Table 1, the results demonstrate that the embedding of EM residual blocks at different levels in the DSD-DETR backbone network significantly impacts model performance. Overall, a reasonable two-layer configuration achieves a good balance between detection accuracy and robustness. For example, when the EM module is placed in layers 6 and 7, the model performs excellently across all metrics, with mAP50 reaching 0.886, clearly demonstrating that deep embedding effectively enhances the model's capture and understanding of the global semantic features of dam surface defects. In contrast, while multi-layer, large-scale embedding (such as layers 4, 5, and 6 or 4, 5, 6, and 7) performs reasonably well on individual metrics, its overall performance is somewhat scattered, possibly due to feature redundancy preventing the model from fully utilizing its discriminative power. Notably, while shallow embedding offers advantages in understanding edge details such as cracks and spalling, it falls slightly short in understanding the global distribution of defects. Comprehensive analysis shows that the reasonable arrangement of EM residual blocks in the deep feature extraction stage can best leverage their advantages, not only improving the accuracy and stability of detection, but also providing more solid technical support for the intelligent identification of complex dam defects.

[0038] Table 1 Comparison of the number of application layers for the EM residual module

[0039] S6. Input the test set images into the trained DSD-DETR model to obtain the detection results.

[0040] The DSD-DETR model proposed in this invention is built upon S2, S3, and S4. After the dataset preprocessing and training hyperparameter configuration tasks in S1 are completed, the training and validation sets are input into the model proposed in this invention for training. After 200 rounds of forward inference, loss calculation, backpropagation, gradient update, and validation set metric evaluation, the model converges to a set of optimal weight parameters, which can be used for various subsequent detection tasks.

[0041] like Figure 4 The figure shows the performance curves of the model during training. After 200 training cycles, the model's positive metrics on the dataset (precision, recall, mAP) are as follows. 50 and mAP 50：95 All of them reached a convergence state. At the same time, the PR curve and F1 confidence curve for each defect type (cracks, spalling, water leakage, exudate and spalling) reached the optimal state.

[0042] Precision rate ( Recall is used to measure the reliability of model predictions, reflecting the proportion of targets correctly identified as diseases out of all those judged as such. It effectively demonstrates the model's ability to avoid false positives. TP (TruePositive) represents the number of true diseases that were correctly detected. FP (False Positive) represents the number of true non-diseases that were incorrectly identified as diseases. This emphasizes the model's sensitivity, measuring the proportion of diseases successfully detected out of all real-world diseases. It's a crucial basis for assessing false negatives; FN (False Negative) represents the number of diseases that are present but not detected. Since precision and recall often involve a trade-off, the F1 score (…) As a harmonic mean of the two, it can provide a more balanced performance evaluation while taking into account both accuracy and completeness.

[0043] mAP ( As a comprehensive indicator in the field of target detection, it comprehensively reflects the overall detection level of the model in the task of identifying multiple types of dam defects by calculating the average accuracy across different categories.

[0044] The detection method proposed in this invention achieves better results in terms of accuracy, recall and mean precision compared with the original RT-DETR-r18 model.

[0045] S7. Analysis and verification.

[0046] To verify the superiority of the DSD-DETR model of this invention, such as Figure 5 As shown in the figure, the image displays a visualized heatmap of the baseline model RT-DETR with the improved module added. Based on the salient region distribution characteristics of the heatmap, DSD-DETR exhibits higher accuracy in detecting surface defects of concrete dams in complex background scenarios. Its higher thermal values ​​allow for more precise focusing on key areas of dam defects. Compared to the baseline model, DSD-DETR, through the synergistic effect of the EM residual block and MSFF multilayer perceptron, not only achieves higher thermal value focusing but also effectively avoids the false detection and missed detection problems present in the baseline model when detecting weaker cracks and more complex defects.

[0047] To verify the superiority of the DSD-DETR model of this invention, an ablation experiment was conducted on the invention, and the performance of each module was compared with the benchmark model. The experimental results are shown in Table 2 below.

[0048] Table 2 Ablation Test Results for Each Module As can be seen, although the EM residual block improves the model's accuracy, it increases the number of parameters. Introducing the MSFF multilayer perceptron concept not only increases the model's detection accuracy but also slightly reduces the number of parameters, saving computational overhead. The proposed DSD-DETR model achieves a 4.4% improvement in mAP50 and an 8.0% reduction in parameters compared to the baseline model on the dam defect dataset. This improves the model's detection performance without increasing the number of parameters.

[0049] Therefore, the present invention employs the above-mentioned method for detecting apparent defects in dams based on improved DSD-DETR, which can significantly improve the accuracy of detecting apparent damage in concrete dams and save a great deal of time, material resources, and financial resources.

[0050] Example 2: This example provides a computer device and a computer-readable storage medium for implementing the dam appearance defect detection method based on improved DSD-DETR described in Example 1.

[0051] The computer device may be a desktop computer, laptop computer, server, edge computing device, UAV ground station equipment, image processing terminal, mobile terminal, or other electronic device with data processing capabilities. The processor may be a central processing unit, graphics processing unit, digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other processor capable of performing computational tasks. The memory may include read-only memory, random access memory, flash memory, solid-state drive (SSD), hard disk drive (HDD), or other read / write storage devices, used to store program instructions, model weight parameters, input images, training sample annotation information, detection results, and related intermediate data.

[0052] The computer-readable storage medium stores a computer program, which, when executed by a processor, can also implement the steps in the dam appearance defect detection method based on improved DSD-DETR described in Embodiment 1. The computer-readable storage medium can be an internal storage unit of a computer device, such as a hard disk, solid-state drive, or memory; or it can be an external storage device, such as a portable hard drive, USB flash drive, memory card, secure digital card, or flash memory card. Furthermore, the computer-readable storage medium may include both internal storage units and external storage devices.

[0053] Those skilled in the art will understand that the aforementioned computer equipment and computer-readable storage media are merely exemplary carriers for implementing the technical solutions of this invention, and their specific hardware forms, storage medium types, and device combinations do not constitute a limitation on the scope of protection of this invention. Any equipment or medium capable of storing and executing corresponding computer programs to implement the dam surface defect detection method described in this invention should fall within the scope of protection of this invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting apparent defects in dams based on an improved DSD-DETR, characterized in that, Includes the following steps: S1. Obtain the dataset of complex apparent defects of the dam, and preprocess the dataset, dividing the preprocessed dataset into training set, validation set and test set. S2. Using the RT-DETR model as the baseline model, replace the residual module of the baseline model with the EM-BasicBlock residual module, and improve the residual module of the RT-DETR model backbone network. S3. Replace the AIFI module of the baseline model with the MSFF module and improve the AIFI module in the RT-DETR model encoder; S4. Replace the loss function of the baseline model with a dynamic equilibrium loss function to improve the loss function of the RT-DETR model; S5. Train the improved DSD-DETR model using the training and validation sets; S6. Input the test set into the trained DSD-DETR model to obtain the detection results; The specific steps of the preprocessing in S1 are as follows: S11. Scale the original image to 640×640 pixels; S12. Perform image enhancement on the scaled original image. Image enhancement includes rotation, flipping, Gaussian blur, contrast adjustment, and brightness adjustment. S13. Label the disease information, and label cracks, peeling, water leakage, exudate and spalling as CR, SP, WL, FBD and EA respectively; S14. After the annotation is completed, divide the dataset into training set, validation set and test set in a ratio of 8:1:1; The specific steps of S3 are as follows: S31. Input a feature tensor The channel dimension is increased through pointwise convolution operations; S32. Orient the unfolded tensor to two parallel branches and enhance the extraction of multi-scale damage information through extended 3×3 and normal 3×3 depth convolutions. S33. Feature enhancement is performed on both branches using learnable weights and bias variables; S34. After passing the inverse fast Fourier transform, the activation output on the branch of the extended 3×3 depth convolution is multiplied pixel by pixel with the direct output of the other branch. S35. Perform a 1×1 convolution on the fused features.

2. The method for detecting dam surface defects based on an improved DSD-DETR according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Two identical fully connected layers are generated by depthwise convolution with a 7×7 kernel. S22. Two fully connected layers are combined into a completely new fully connected layer through a star operation; S23. Reduce computational load by using 1×1 convolutional compression channels to achieve feature extraction.

3. The method for detecting dam surface defects based on an improved DSD-DETR according to claim 2, characterized in that, The specific formula for the star operation in S22 is as follows: ； in, d is the input channel number, which is expanded to accommodate multiple output channels. Simultaneously processing with Multiple feature elements.

4. The method for detecting dam surface defects based on improved DSD-DETR according to claim 1, characterized in that: The EM-BasicBlock residual module in S2 only replaces the 6th and 7th layers of the RT-DETR model backbone network, and determines its optimal embedding depth and number by comparing the detection accuracy, recall rate and mAP index under different layer configurations.

5. The method for detecting apparent defects in dams based on an improved DSD-DETR according to claim 1, characterized in that, The dynamic equilibrium loss function in S4 is the joint loss function of Focaler-IoU and CIoU, Focaler-CIoU. The specific formula for Focaler-CIoU is as follows: ； in, Focaler-CIoU is the joint loss function used for training optimization of bounding box regression. For complete intersection-over-union, it is used to comprehensively measure the degree of overlap between the predicted box and the ground truth box, the distance between the center points, and the consistency of the aspect ratio; Intersection over Union (IoU) is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box. This is the Focaler-IoU term, used to dynamically reweight the regression gradients of samples of different quality.

Images