A dam underwater disease identification method, system and medium

Abstract

The application discloses a dam underwater disease identification method and system and a medium, and belongs to the technical field of dam underwater disease intelligent identification. The method comprises the following steps: acquiring a dam underwater disease image, inputting a pre-constructed disease classification model, and outputting a classification result; the construction of the disease classification model comprises the following steps: taking UniRepLKNet as a basic network, replacing the SmaK module in the UniRepLKNet with an efficient feature extraction module, replacing the Stem module in the UniRepLKNet with a detail-preserving Stem module DPStem, and replacing the expansion reparameterization structure in the LarK module with an Offset convolution structure, so as to obtain an identification model. The application can effectively cope with underwater environmental noise interference, image blur, insufficient illumination, color distortion, and the problems that a disease target is fine and irregular in shape, and can enhance the representation ability of local detail features, multi-scale disease features and long-distance dependent information, so as to improve the accuracy, stability and reliability of identification.

Description

Technical Field

[0001] This invention relates to the field of underwater dam defect identification technology, and in particular to a method, system and medium for underwater dam defect identification. Background Technology

[0002] In modern society, hydropower dams serve as crucial infrastructure for flood control, water supply, irrigation, navigation, and hydropower generation. Their safe and stable operation is of paramount importance to socio-economic development and energy supply. However, dams operate under complex environments for extended periods, subjected to the combined effects of water pressure, temperature variations, material aging, and external loads. This can lead to various defects on the dam surface and in localized structural areas, such as cracks, spalling, voids, and exposed rebar. These defects not only weaken the overall integrity and durability of the dam structure but may also expand, impacting its long-term operational safety. Therefore, timely and accurate identification of dam defects is essential for ensuring the safe operation of hydropower projects.

[0003] Traditional dam defect identification mainly relies on manual inspections or visual checks. These methods suffer from low efficiency, high subjectivity, insufficient automation, and high operational safety risks, especially in complex underwater environments where these shortcomings are more pronounced. Underwater scenarios generally present challenges such as poor lighting conditions, low image resolution, color distortion, strong noise interference, and subtle target features, making accurate extraction and identification of defects difficult. Traditional methods cannot meet the demands for high-precision, intelligent underwater dam defect detection.

[0004] In recent years, with the development of computer vision and deep learning technologies, convolutional neural networks (CNNs) have demonstrated strong feature learning and generalization capabilities in tasks such as image classification, object detection, and semantic analysis, gradually becoming an important technical means for image recognition in complex scenes. However, existing CNNs still suffer from problems such as high computational cost, insufficient long-distance dependency information capture, and limited ability to represent fine-grained defect features when processing large-sized images or complex underwater scene images, resulting in the need to improve the accuracy of underwater dam defect identification. Therefore, developing a method for identifying underwater dam defects that can adapt to complex underwater environments, balance local detail feature extraction and global context modeling capabilities, and possess high recognition accuracy and stability has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system and medium for underwater dam defect identification. This method can effectively address problems such as noise interference, image blurring, insufficient lighting, color distortion and fine and irregular shapes of defect targets in the underwater environment, thereby improving the accuracy, stability and reliability of underwater dam defect image identification and providing reliable technical support for underwater dam safety monitoring and operation and maintenance.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution: On the one hand, the present invention provides a method for identifying underwater dam defects, including: Obtain raw underwater images of dam defects; The original underwater dam defect images are input into a pre-built defect identification model, and the underwater dam defect identification results are output. The processing steps of the disease identification model include: The DStem module is used to downsample and extract initial features from the original underwater dam defect image to obtain an output feature map. The output feature map is subjected to hierarchical feature extraction through multiple stages to obtain local features and high-level semantic features at different levels. The feature extraction process employs an efficient feature extraction module to model local texture and fine-grained disease features, and an improved LarK module containing an Offset convolution structure to model long-distance dependency information and global context information. Global average pooling is performed on the final output features to obtain a one-dimensional feature vector; The one-dimensional feature vector is input into the fully connected layer, which outputs the predicted probability corresponding to each disease category, and the disease identification result is determined based on the predicted probability.

[0007] Optionally, the efficient feature extraction module includes: Perform layer normalization on the input features; Extracting local features using depthwise convolution; After inputting the local features into two fully connected layers, element-wise multiplication aggregation of features from different channels is performed through star operations to enhance the nonlinear relationship between features; The aggregated features are further enhanced by shifting the input to the Shift Module and performing depthwise convolution. The enhanced features are concatenated and fused with the aforementioned branch features, and the number of channels is restored through pointwise convolution to obtain the output features.

[0008] Optionally, the Shift Module shift operation includes: The input feature map is segmented along the channel dimension; Perform a horizontal shift operation on the segmented channel; Perform a vertical shift operation on the other part of the segmented channel; The features of each channel after shifting are fused to enhance the interaction between different regions of the feature map and the ability to represent disease details.

[0009] Optionally, the DStem module includes: Perform a 1×1 convolution on the input feature map to obtain the initial feature map; The initial feature map is spatially reduced by using a 3×3 depthwise convolution branch and a PUDown downsampling branch respectively. The outputs of the two branches are then concatenated and fused, and the number of channels is restored by a 1×1 convolution. The spatial resolution of the feature map is further reduced by using a 3×3 depthwise convolution branch and a max pooling branch respectively. The outputs of the two branches are then concatenated and fused, and local features are extracted by 1×1 convolution. Multi-scale feature extraction of fused features is performed using 1×1 convolution, 3×3 convolution, and 5×5 convolution. Multi-scale features are spliced ​​and fused, and the number of channels is restored by 1×1 convolution to obtain the output features of the DPStem module.

[0010] Optionally, the PUDown downsampling branch includes: The input feature map is pixel rearranged to decompose it into multiple sub-feature maps; The multiple sub-feature maps are concatenated along the channel dimension; Perform a 1×1 convolution on the stitched feature map to restore the number of channels, and output the downsampled feature map; The PUDown downsampling branch reduces resolution by rearranging the spatial relationships between pixels in order to preserve detailed information in the diseased image during the downsampling process.

[0011] Optionally, the improved LarK module includes: Local texture and edge details of the input features are modeled using 3×3 convolutions and 3×3 convolutions with a dilation of 2. The input features are modeled using an Offset convolutional structure to perform long-distance dependency modeling. The Offset convolutional structure utilizes multiple 3×3 convolutional branches, channel rearrangement, and spatial shifting operations to obtain a receptive field equivalent to a large-size convolutional kernel. Multi-branch convolutional structures are reparameterized and fused during the inference phase to obtain a single-path convolutional structure output.

[0012] Optionally, the Offset convolutional structure includes: Each channel of the input feature map is convolved using multiple 3×3 convolution kernels to obtain intermediate features; The intermediate features are channel rearranged, and the rearranged features are divided into multiple groups; Perform a spatial shift operation with a preset direction and preset offset for each group of features; The shifted features are summed and aggregated to obtain the path output. The outputs of multiple paths are aggregated element by element and then normalized to obtain the output of the Offset convolution structure. The Offset convolutional structure achieves feature diffusion through spatial shifting and multi-branch fusion to enhance cross-regional propagation capabilities, global context aggregation capabilities, and long-distance dependency modeling capabilities.

[0013] Optionally, the training of the disease identification model includes: Acquire underwater image data of dam defects formed by extracting video frames collected by an underwater robot; The image data undergoes data augmentation processing, including cropping and flipping. The enhanced image data is divided into a training set and a test set; The disease identification model is trained using the AdamW optimizer to obtain a trained disease identification model.

[0014] Optionally, the underwater images of dam defects include one or more types of defects such as cracks, spalling, holes, and exposed rebar.

[0015] Secondly, the present invention provides an underwater dam defect identification system, comprising: The image acquisition module is used to acquire raw underwater images of dam defects; The disease identification module is used to input the original underwater disease images of the dam into a pre-built disease identification model and output the underwater disease identification results of the dam. The model building module is used to build the disease identification model; The model building module is used to obtain the UniRepLKNet base network, replace the SmaK module in the UniRepLKNet base network with an efficient feature extraction module, replace the Stem module with the detail-preserving Stem module DPStem, and replace the dilated reparameterization structure in the LarK module with the Offset convolution structure to obtain the constructed disease identification model.

[0016] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the underwater dam defect identification method described in the first aspect.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: Compared with existing technologies, this invention improves the UniRepLKNet network structure, effectively addressing issues such as high noise interference, low resolution, and subtle disease features in underwater environments, thereby enhancing the accuracy and stability of underwater dam disease image recognition. Specifically, the efficient feature extraction module enhances local feature extraction capabilities and model generalization performance, the DStem module improves the preservation of detailed information and multi-scale features, and the improved LarK module enhances the receptive field and global context modeling capabilities by introducing Offset convolution. This achieves superior disease recognition results with lower computational overhead and demonstrates significant engineering application value. Attached Figure Description

[0018] Figure 1 The diagram shown is a flowchart of one embodiment of the underwater dam defect identification method of the present invention. Figure 2 The diagram shown is a structural schematic of the high-efficiency feature extraction module in one embodiment of the present invention; Figure 3 The diagram shown is a structural schematic of the shift module shift operation in one embodiment of the present invention; Figure 4 The diagram shown is a structural schematic of the PUDown module in one embodiment of the present invention; Figure 5 The diagram shown is a structural schematic of the DStem module in one embodiment of the present invention; Figure 6 The diagram shown is a structural schematic of the improved LarK module in one embodiment of the present invention; Figure 7 The diagram shown is a structural schematic of the Offset convolution module in one embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0020] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Example 1

[0021] like Figure 1As shown in this embodiment, an underwater dam defect identification method is introduced. This method can solve the problems of low identification accuracy, easy loss of detailed information, insufficient long-distance dependency modeling, and limited computational efficiency of traditional underwater dam defect identification methods in complex underwater environments. It can effectively cope with the problems of low resolution, large noise interference, and subtle and irregular defect features in underwater environments, thereby improving the accuracy, stability and robustness of identifying defects such as cracks, spalling, holes and exposed rebar, and providing reliable technical support for the safety monitoring and maintenance of underwater dams.

[0022] The method includes the following steps: Step 1: Obtain raw underwater images of dam defects, specifically: Underwater robots were used to collect underwater video data in a real-world hydroelectric dam setting. The collected videos were then processed by frame extraction to obtain raw underwater images of dam defects. These images were then filtered and organized to construct a dataset of underwater dam defect images. To improve data diversity and model generalization ability, the raw images underwent further data augmentation processing, including cropping and flipping.

[0023] In this embodiment, the self-built underwater dam defect dataset contains 2824 single defect image samples, covering four typical defects: cracks, spalling, voids, and exposed rebar. Among them, there are 1953 crack samples, 315 spalling samples, 113 void samples, and 443 exposed rebar samples. The dataset is divided into 80% training set and 20% test set to balance the independence of model training effect and test evaluation.

[0024] Step 2: Input the original underwater dam defect images into the pre-built defect identification model, and output the underwater dam defect identification results, specifically: The disease identification model is constructed based on UniRepLKNet, with the SmaK module in the original network replaced by a high-efficiency feature extraction module. A schematic diagram of the high-efficiency feature extraction module is shown below. Figure 2 As shown, the Stem module is replaced with the detail-preserving Stem module DStem, and the dilated reparameterization structure in the LarK module is replaced with an Offset convolution structure to obtain the constructed disease identification model. The model first uses the DStem module to downsample the input image and extract initial features, then performs hierarchical feature learning through four stages, and finally uses global average pooling and fully connected layers to output the predicted probabilities of each disease category, determining the final identification result based on the principle of maximizing probability.

[0025] Furthermore, in this embodiment, the number of output channels for the four stages are set to 48, 96, 192, and 384, respectively. As the network depth increases, the spatial resolution of the feature map gradually decreases, while the semantic expressive power gradually increases, thereby achieving a step-by-step abstraction from shallow local details to deep high-level semantic information. Specifically, the efficient feature extraction module enhances the extraction capabilities of local textures, edges, and fine-grained disease features; the improved LarK module, which includes an offset convolution structure, enhances the modeling capabilities of long-distance dependency information and global contextual information. Finally, the output feature maps are aggregated through global average pooling, mapping high-dimensional features to one-dimensional feature vectors, which are then input into a fully connected layer to obtain the prediction results corresponding to cracks, spalling, holes, and exposed rebar.

[0026] Step 3: Perform local feature enhancement using an efficient feature extraction module, specifically: To address the limitations of the SmaK module in the original UniRepLKNet in representing local details and generalization ability of underwater disease images, this embodiment designs an efficient feature extraction module. This module first performs layer normalization on the input features to stabilize the training process; then, it uses deep convolution to extract local features; next, the features are input into two fully connected layers, and element-wise multiplication is used to aggregate features from different channels through a star-shaped operation to enhance the nonlinear relationships between features and the high-dimensional representation capability; then, the Shift Module shift operation and deep convolution are used to further enhance the spatial information interaction capability; finally, the enhanced features are concatenated and fused with the aforementioned branch features, and the number of channels is restored through pointwise convolution to obtain the output features. This module can improve local feature extraction capability, nonlinear expression capability, and gradient flow performance without significantly increasing the number of parameters.

[0027] Furthermore, such as Figure 3 As shown, the Shift Module shifting operation includes: segmenting the input feature map along the channel dimension, performing a horizontal shift operation on a portion of the channels, performing a vertical shift operation on another portion of the channels, and fusing the features of the shifted channels. This shifting operation enhances the interaction between different regions of the feature map and improves the sensitivity to minute and irregular disease features such as cracks and holes.

[0028] Step Four: As Figure 5 As shown, detail-preserving downsampling is performed using the DStem module, specifically as follows: To address the issue of detail loss during downsampling in the original Stem module, this embodiment designs the DStem module. This module first extracts initial features from the input image using a 1×1 convolution. Then, it performs a first spatial resolution reduction on the feature map using a 3×3 depthwise convolution branch and a PUDown downsampling branch, concatenating the outputs of the two branches and restoring the channel count using a 1×1 convolution. Next, it further reduces the spatial resolution using a 3×3 depthwise convolution branch and a max-pooling branch, concatenating the outputs of the two branches and extracting local features using a 1×1 convolution. Finally, it performs multi-scale feature extraction using 1×1, 3×3, and 5×5 convolutions, concatenating and fusing the multi-scale features and restoring the channel count using a 1×1 convolution to obtain the output features of the DStem module. Through multi-branch progressive downsampling and multi-scale convolutional fusion, the DStem module can effectively preserve the detail and semantic information in the diseased image while reducing resolution.

[0029] Furthermore, such as Figure 4 As shown, the PUDown downsampling branch decomposes the input feature map into multiple sub-feature maps by rearranging pixels, then concatenates these sub-feature maps along the channel dimension, and finally restores the number of channels through a 1×1 convolution to output the downsampled feature map. Compared with traditional pooling or direct convolution downsampling, this method achieves resolution reduction by rearranging the spatial relationships between pixels, which helps to retain more high-level features and detailed information during the downsampling process.

[0030] Step 5: As Figure 6 As shown, long-distance dependency modeling is performed using an improved LarK module, specifically as follows: To address the shortcomings of large-kernel convolutions in detail preservation and long-range dependency modeling, this embodiment improves the LarK module by introducing an Offset convolution structure. The improved LarK module first uses 3×3 convolutions and 3×3 convolutions with a dilation of 2 to model local texture and edge details of the input features; subsequently, it employs an Offset convolution structure to model long-range dependencies of the input features. This Offset convolution structure, through multiple 3×3 convolution branches, channel rearrangement, and spatial shifting operations, achieves a receptive field equivalent to a large-size convolution kernel with lower computational cost, thereby enhancing the representation ability of sparse structural features, cross-regional correlation information, and global contextual information. During the inference phase, the multi-branch convolution structure can be further fused into a single-path convolution structure through reparameterization to balance recognition accuracy and inference efficiency.

[0031] Furthermore, such as Figure 7As shown, the Offset convolutional structure includes: convolving each channel of the input feature map with multiple 3×3 convolutional kernels to obtain intermediate features; rearranging the channels of the intermediate features and dividing the rearranged features into multiple groups; performing a spatial shift operation with a preset direction and preset offset on each group of features; summing and aggregating the shifted groups of features to obtain path outputs; then adding and aggregating the multiple path outputs element-wise and normalizing them to obtain the Offset convolutional structure output. Through the above operations, information from the same input location can play a role simultaneously in multiple spatial locations and multiple channels, thereby achieving feature diffusion and improving the ability to model long-distance dependencies and aggregate global context.

[0032] Step Six: Train and test the disease identification model, specifically as follows: This embodiment was conducted on an Ubuntu 22.04 operating system. The deep learning framework used was PyTorch 1.13.0, the programming language was Python 3.8, and the hardware platform included an Intel Core i5-10400F processor, an Nvidia GeForce RTX 3090 graphics card, and 32GB of RAM. During training, the batch size was set to 8, the initial learning rate was set to 0.001, and the optimizer was AdamW. The number of training epochs on the underwater dam disease dataset was set to 100.

[0033] In this embodiment, ablation experiments were conducted to verify the contributions of the efficient feature extraction module, the DStem module, and the improved LarK module to the improvement of model performance. When all three modules were introduced simultaneously, the model had 6.69M parameters, 1.04G FLOPS, and a Top-1 accuracy of 98.39%, which is 1.97% higher than the baseline UniRepLKNet and outperforms the comparison model. This indicates that the method in this embodiment can effectively improve the accuracy and robustness of underwater dam defect image recognition while maintaining low computational overhead. Example 2

[0034] Based on Example 1, this example introduces an underwater dam defect identification system, including: The image acquisition module is used to acquire raw underwater images of dam defects. The defect identification module is used to input the original underwater dam defect images into a pre-built defect identification model and output the underwater dam defect identification results. The model building module is used to build the disease identification model. The model building module is used to obtain the UniRepLKNet basic network, replace the SmaK module in the UniRepLKNet basic network with an efficient feature extraction module, replace the Stem module with the detail-preserving Stem module DPStem, and replace the dilated reparameterization structure in the LarK module with the Offset convolution structure to obtain the constructed disease identification model.

[0035] The specific functions of each module described above are explained in the relevant content of Embodiment 1 or 2, and will not be repeated here. Example 3

[0036] This embodiment introduces a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the underwater dam defect identification method described in Embodiment 1.

[0037] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0038] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0039] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0040] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0041] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for identifying underwater defects in dams, characterized in that, include: Obtain raw underwater images of dam defects; The original underwater dam defect images are input into a pre-built defect identification model, and the underwater dam defect identification results are output. The construction of the disease identification model includes: Obtain the UniRepLKNet basic network; The SmaK module in the UniRepLKNet basic network is replaced with an efficient feature extraction module, the Stem module is replaced with a detail-preserving Stem module DStem, and the dilated reparameterization structure in the LarK module is replaced with an Offset convolution structure to obtain the constructed disease identification model. The high-efficiency feature extraction module includes: Perform layer normalization on the input features; Extracting local features using depthwise convolution; After inputting the local features into two fully connected layers, element-wise multiplication aggregation of features from different channels is performed through star operations to enhance the nonlinear relationship between features; The aggregated features are passed through a fully connected layer to obtain intermediate features; The intermediate features are further enhanced by shifting the input to the Shift Module and performing a depthwise convolution. The enhanced features are concatenated and fused with the input features, the intermediate features, and the features after the shift operation, and the number of channels is restored by pointwise convolution to obtain the output features.

2. The method for identifying underwater defects in dams according to claim 1, characterized in that, The processing steps of the disease identification model include: The DStem module is used to downsample and extract initial features from the original underwater dam defect image to obtain an output feature map. The output feature map is subjected to hierarchical feature extraction in four stages to obtain local features and high-level semantic features at different levels. Among them, the feature map uses an efficient feature extraction module to model local texture and fine-grained disease features, and an improved LarK module with an offset convolution structure to model long-distance dependency information and global context information. Global average pooling is performed on the final output features to obtain a one-dimensional feature vector; The one-dimensional feature vector is input into the fully connected layer, which outputs the predicted probability corresponding to each disease category, and the disease identification result is determined based on the predicted probability.

3. The method for identifying underwater defects in dams according to claim 1, characterized in that, The shift module shift operation includes: The input feature map is segmented along the channel dimension; Perform a horizontal shift operation on the segmented channel; Perform a vertical shift operation on the other part of the segmented channel; The features of each channel after shifting are fused to enhance the interaction between different regions of the feature map and the ability to represent disease details.

4. The method for identifying underwater defects in dams according to claim 2, characterized in that, The DStem module includes: Perform a 1×1 convolution on the input feature map to obtain the initial feature map; The initial feature map is spatially reduced by using a 3×3 depth convolution branch and a PUDown downsampling branch respectively. The outputs of the two branches are then concatenated and fused, and the number of channels is restored by a 1×1 convolution. The spatial resolution of the feature map is further reduced by using a 3×3 depthwise convolution branch and a max pooling branch respectively. The outputs of the two branches are then concatenated and fused, and local features are extracted by 1×1 convolution. 1×1 convolution, 3×3 convolution and 5×5 convolution are used to extract local features at multiple scales; Multi-scale features are spliced ​​and fused, and the number of channels is restored by 1×1 convolution to obtain the output features of the DPStem module.

5. The method for identifying underwater defects in dams according to claim 4, characterized in that, The PUDown downsampling branch includes: The input feature map is pixel rearranged to decompose it into multiple sub-feature maps; The multiple sub-feature maps are concatenated along the channel dimension; Perform a 1×1 convolution on the stitched feature map to restore the number of channels, and output the downsampled feature map; The PUDown downsampling branch reduces resolution by rearranging the spatial relationships between pixels in order to preserve detailed information in the diseased image during the downsampling process.

6. The method for identifying underwater defects in dams according to claim 2, characterized in that, The improved LarK module includes: Local texture and edge details of the input features are modeled using 3×3 convolutions and 3×3 convolutions with a dilation of 2. The input features are modeled using an Offset convolutional structure to perform long-distance dependency modeling. The Offset convolutional structure utilizes multiple 3×3 convolutional branches, channel rearrangement, and spatial shifting operations to obtain a receptive field equivalent to a large-size convolutional kernel. Multi-branch convolutional structures are reparameterized and fused during the inference phase to obtain a single-path convolutional structure output.

7. The method for identifying underwater defects in dams according to claim 6, characterized in that, The Offset convolutional structure includes: Each channel of the input feature map is convolved using multiple 3×3 convolution kernels to obtain intermediate features; The intermediate features are channel rearranged, and the rearranged features are divided into multiple groups; Perform a spatial shift operation with a preset direction and preset offset for each group of features; The shifted features are summed and aggregated to obtain the path output. The outputs of multiple paths are aggregated element by element and then normalized to obtain the output of the Offset convolution structure. The Offset convolutional structure achieves feature diffusion through spatial shifting and multi-branch fusion to enhance cross-regional propagation capabilities, global context aggregation capabilities, and long-distance dependency modeling capabilities.

8. A system for identifying underwater defects in dams, characterized in that, include: The image acquisition module is used to acquire raw underwater images of dam defects; The disease identification module is used to input the original underwater disease images of the dam into a pre-built disease identification model and output the underwater disease identification results of the dam. The model building module is used to build the disease identification model; The model building module is used to obtain the UniRepLKNet basic network, replace the SmaK module in the UniRepLKNet basic network with an efficient feature extraction module, replace the Stem module with the detail-preserving Stem module DStem, and replace the dilated reparameterization structure in the LarK module with the Offset convolution structure to obtain the constructed disease identification model. The high-efficiency feature extraction module includes: Perform layer normalization on the input features; Extracting local features using depthwise convolution; After inputting the local features into two fully connected layers, element-wise multiplication aggregation of features from different channels is performed through star operations to enhance the nonlinear relationship between features; The aggregated features are passed through a fully connected layer to obtain intermediate features; The intermediate features are further enhanced by shifting the input to the Shift Module and performing a depthwise convolution. The enhanced features are concatenated and fused with the input features, the intermediate features, and the features after the shift operation, and the number of channels is restored by pointwise convolution to obtain the output features; The training of the disease identification model includes: acquiring underwater disease image data of the dam formed by extracting video frames collected by an underwater robot; performing data augmentation processing such as cropping and flipping on the image data; dividing the augmented image data into a training set and a test set; and training the disease identification model using the AdamW optimizer to obtain the trained disease identification model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the underwater dam defect identification method according to any one of claims 1 to 7.

Images