An underwater edge-guided dual-stream concrete defect detection system and method

Abstract

The present application relates to the technical field of underwater image processing, and particularly relates to a kind of underwater edge guided double-flow concrete defect detection system and method.The system contains input, double-flow feature extraction, hierarchical double-flow fusion, cross-scale fusion and detection head module.Input module obtains RGB image and generates corresponding edge contour map;Double-flow feature extraction module extracts multi-scale features in parallel with RGB semantic branch and edge geometry branch, each branch is built-in lightweight enhancement module;Hierarchical double-flow fusion module differentiates processing for middle and high layer features, middle layer realizes edge guided enhancement and adaptive weighted fusion, high layer completes multi-flow information interaction and feature calibration;Cross-scale fusion module integrates multi-level features, and outputs defect category, confidence and bounding box through decoupling detection head.The present application effectively solves the problems of underwater defect boundary blur, low detection accuracy and model difficult deployment, realizes high-precision lightweight automatic detection, and adapts to underwater robot edge device application.

Description

Technical Field

[0001] This invention relates to the field of underwater image processing technology, and in particular to an underwater edge-guided dual-flow concrete defect detection system and method. Background Technology

[0002] Underwater concrete structures in water-related projects such as bridges, dams, and ports are constantly exposed to complex aquatic environments characterized by high humidity, high water pressure, and chemical corrosion. This makes them prone to various defects such as cracks, rebar corrosion, exposed rebar, and concrete spalling. These defects continuously deteriorate the structure's load-bearing capacity and durability, and in severe cases, can lead to major safety accidents such as bridge collapses and dam leaks. Therefore, conducting regular and accurate defect detection of underwater concrete structures is a core element in ensuring the safe operation of these projects.

[0003] Current underwater concrete defect detection technologies mainly employ deep learning-based visual detection methods, which are currently the mainstream solutions. However, these methods have several drawbacks: Firstly, underwater water scattering, turbidity, and non-uniform lighting severely weaken the texture contrast of defect areas, leading to significant blurring of the boundary features of slender defects such as cracks. Detection models relying solely on single RGB semantic features cannot simultaneously ensure reliable defect identification and accurate localization, resulting in poor boundary localization stability. Secondly, crack-type defects are slender and topologically continuous, while spalling and corrosion defects are planar and irregular in shape. A single feature branch cannot simultaneously meet the dual requirements of semantic discrimination and geometric localization for different defect types, easily leading to missed detection of small defects and misalignment of large defects. Finally, in practical engineering, existing multimodal fusion detection methods generally employ complex fusion architectures to improve accuracy, resulting in a dramatic increase in model parameters and computational load, making them unsuitable for the real-time detection needs of edge devices such as underwater robots. Meanwhile, lightweight single-stream models cannot solve the problems of boundary blurring and feature degradation in underwater scenes, making it difficult to meet engineering requirements in terms of detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide an underwater edge-guided dual-flow concrete defect detection system and method to solve the technical problems faced by existing underwater concrete defect detection technologies, such as difficulty in locating fuzzy boundaries, poor adaptability to multiple defect morphologies, and the inability to balance detection accuracy and deployment efficiency, thereby achieving high-precision, lightweight, and automated detection of underwater concrete defects.

[0005] The technical solution of this invention is: an underwater edge-guided dual-flow concrete defect detection system, comprising: The input module is used to acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map; The dual-stream feature extraction module is used to extract features from the RGB image and the edge contour map through the RGB semantic branch and the edge geometry branch respectively, and outputs the RGB multi-scale feature map and the edge multi-scale feature map. The hierarchical dual-stream fusion module is used to perform differential fusion of features at different levels in the RGB multi-scale feature map and the edge multi-scale feature map. This includes: generating attention information based on the mid-level edge features and applying it to the mid-level RGB features using residual modulation to achieve edge-guided enhancement; obtaining a first fusion feature through adaptive weighted fusion; refining the first fusion feature to obtain a mid-level fusion feature; and fusing the high-level RGB features and high-level edge features based on bidirectional attention multi-stream information interaction to obtain a second fusion feature; modulating the second fusion feature using spatial weights generated from the second fusion feature to achieve feature calibration, thus obtaining a high-level fusion feature. The cross-scale fusion module is used to fuse mid-level fusion features, high-level fusion features, and RGB low-level features to output multi-scale detection features; The detection head module is used to output the defect category, confidence level, and bounding box coordinates of the defect based on multi-scale detection features, through classification and regression branches.

[0006] Preferably, both the RGB semantic branch and the edge geometry branch contain multiple levels of feature extraction structures, and each level outputs a feature map of the corresponding scale. In the RGB semantic branch, each level of the feature extraction structure is equipped with a CDMS module for multi-scale feature extraction and a CTAM module for channel response adjustment. In the edge geometry branch, each level of the feature extraction structure is equipped with an EEDC module for preserving edge information and an EAEM module for spatial feature enhancement.

[0007] Preferably, the CDMS module includes a parallel four-branch structure, namely a micro-scale branch, a local branch, a meso-scale branch, and a macro-scale branch. The outputs of the four branches are spliced ​​along the channel dimension and then subjected to batch normalization and activation functions to obtain RGB multi-scale features. The four branches employ convolutional kernels of different sizes to extract defect features under different receptive fields. Preferably, the EEDC module includes a feature extraction branch and an edge preservation branch; The feature extraction branch uses convolutional layers to downsample the input features to obtain basic semantic features; The edge-preserving branch processes the input features sequentially through a feature transformation layer and a depthwise separable convolutional layer to preserve edge response information during downsampling.

[0008] Preferably, in the layered dual-stream fusion module, the processing of RGB mid-layer features and edge mid-layer features includes: The edge mid-layer features are transformed to generate an edge attention map. The RGB mid-layer features are weighted based on the edge attention map. The weighted result is added to the RGB mid-layer features element by element to obtain the edge guidance enhancement map. The edge guidance enhancement map and the mid-layer edge features are processed by depthwise separable convolutional blocks, and then added element-wise with the mid-layer RGB features and the mid-layer edge features respectively, and then adaptively weighted and fused to obtain the first fused feature. The first fusion feature is refined using a three-dimensional joint attention mechanism to generate a three-dimensional joint attention, including channel attention weights, height direction attention weights, and width direction attention weights. The three-dimensional joint attention is then weighted element-wise with the initial fusion feature to obtain the intermediate fusion feature.

[0009] Preferably, the first fusion feature The generation process is as follows: ; in, and These are RGB adaptive fusion weights and edge adaptive fusion weights, respectively. For depthwise separable convolutional blocks; Enhanced edge-guided graph; This is a mid-layer feature of RGB; This represents the features of the middle layer at the edge.

[0010] Preferably, in the layered dual-stream fusion module, the processing of RGB high-level features and edge high-level features includes: Query vectors, key vectors, and value vectors are constructed based on RGB high-level features and edge high-level features. RGB interaction features and edge interaction features are obtained through multi-stream information interaction based on bidirectional attention. The RGB interaction features and the edge interaction features are weighted and fused to obtain the second fused feature; An edge saliency map is generated based on the second fusion feature. The edge saliency map is then concatenated with the second fusion feature channel. The channel concatenation result is processed to obtain a spatial feature calibration weight. The spatial feature calibration weight is used to calibrate the second fusion feature to obtain a high-level fusion feature.

[0011] Preferably, the feature calibration includes at least two parallel calibration branches constructed based on the second fused feature: The first calibration branch is used to fuse the edge saliency map and spatial feature calibration weights, and to enhance the second fused feature to obtain the first feature calibration map. The second calibration branch is used to perform weighted modulation of the second fused feature based on the spatial feature calibration weight to obtain the second feature calibration map; The first and second calibration maps are fused together, and high-level fusion features are generated based on residual connections.

[0012] Preferably, the feature calibration is implemented by: The first calibration branch is constructed based on the edge saliency map, spatial feature calibration weights, and the second fusion feature: , in, This is the calibration map of the first feature. This is the second fusion feature. The edge enhancement coefficient, For the edge saliency map, The weights are calibrated for the spatial features; A second calibration branch is constructed based on spatial feature calibration weights and second fusion features: , This is the calibration map for the second feature; The first and second feature calibration maps are fused and then processed by convolution to obtain high-level fused features. : , in, It is a 1×1 convolution.

[0013] On the other hand, this application also discloses an underwater edge-guided dual-flow concrete defect detection method, based on any one of the underwater edge-guided dual-flow concrete defect detection systems described in the previous application, comprising: Acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map; Feature extraction is performed on the RGB image and the edge contour map by using the RGB semantic branch and the edge geometry branch respectively, and the output RGB multi-scale feature map and edge multi-scale feature map are generated. Differential fusion is performed on features at different levels in the RGB multi-scale feature map and the edge multi-scale feature map, including: generating attention information based on the mid-level edge features and applying it to the mid-level RGB features in a residual modulation manner to achieve edge guidance enhancement, and obtaining a first fused feature through adaptive weighted fusion; refining the first fused feature to obtain a mid-level fused feature; fusing the high-level RGB features and the high-level edge features based on bidirectional attention multi-stream information interaction to obtain a second fused feature; modulating the second fused feature based on the spatial weights generated by the second fused feature to achieve feature calibration, and obtaining a high-level fused feature; The mid-level fusion features, high-level fusion features, and RGB low-level features are fused to output multi-scale detection features; Based on multi-scale detection features, the system outputs the defect category, confidence level, and bounding box coordinates of the defect through classification and regression branches.

[0014] Compared with the prior art, the advantages of the present invention are: (1) This invention effectively alleviates the problems of boundary blurring and feature instability caused by underwater scattering, turbidity and uneven illumination by using parallel modeling of RGB semantics and edge geometry. On the self-built underwater concrete defect dataset, it achieves 76.3% mAP50 and 54.7% mAP50-95. Compared with the mainstream YOLO series single-flow model, the mAP50 is improved by more than 4 percentage points, and the detection accuracy of boundary-sensitive defects such as cracks and exposed reinforcement is particularly significant.

[0015] (2) The layered dual-stream fusion strategy of the present invention designs a differentiated fusion method based on the characteristics of the middle layer and the high layer. The middle layer fusion focuses on the fine positioning of weak boundary defects, and the high layer fusion strengthens the accurate recognition of global semantics. It can simultaneously adapt to a variety of defects with huge morphological differences, such as slender cracks, surface peeling or corrosion, and achieve stable detection effect on all four types of defects without obvious performance shortcomings.

[0016] (3) The total number of network parameters of the present invention is only 8.6M and the computation amount is only 20.2GFLOPs. While achieving high-precision detection, it maintains extremely low computational complexity. Compared with the existing high-precision multimodal fusion model, the number of parameters and computation amount are reduced by more than one order of magnitude. It has excellent edge deployment potential and can be adapted to the real-time detection needs of embedded devices such as underwater robots. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the underwater edge-guided dual-flow concrete defect detection system according to the present invention; Figure 2 This is a schematic diagram of the structure of the CDMS module described in this invention; Figure 3 This is a schematic diagram of the CTAM module described in this invention; Figure 4 This is a schematic diagram of the EEDC module described in this invention; Figure 5 This is a schematic diagram of the structure of the EAEM module described in this invention; Figure 6 This is a schematic diagram of the edge guidance enhancement and adaptive weighted fusion described in this invention; Figure 7 This is a schematic diagram illustrating the refinement of the features described in this invention; Figure 8 This is a schematic diagram of the multi-stream information interaction described in this invention; Figure 9 This is a schematic diagram of the feature calibration described in this invention; Figure 10 This is a performance comparison chart between the present invention and existing mainstream detection methods. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments: This invention is primarily applied to intelligent defect detection scenarios in underwater concrete structures, including automated inspection of appearance defects in underwater concrete structures such as reservoir dams, bridge piers, port terminals, and cross-sea tunnels. In actual underwater inspection environments, issues such as turbid water scattering, uneven lighting, and low image contrast can severely weaken the boundary features of defects such as concrete cracks, exposed reinforcement, spalling, and corrosion, and make semantic information easily confused by environmental interference. Existing defect detection methods often employ single-branch feature extraction networks or homogeneous fusion strategies for features at different levels, which easily leads to the loss of fine-grained edges and ambiguity in high-level semantic discrimination. It is difficult to simultaneously ensure the accuracy of defect boundary positioning and category recognition, and the feature adaptability for underwater weak-edge defects is poor, failing to meet the high-precision and high-reliability intelligent detection requirements of underwater engineering.

[0019] To address these issues, this invention constructs a dual-stream parallel feature extraction and hierarchical differentiated feature fusion architecture. This architecture focuses on solving the core technical problems of underwater concrete defect boundaries being easily lost, defect semantics being easily confused, and insufficient detection accuracy due to the single feature fusion method at different levels. The technical solution is described in detail below with specific embodiments.

[0020] like Figure 1 As shown, an underwater edge-guided dual-stream concrete defect detection system is constructed based on the Underwater Edge-Guided Dual-stream Defect Network (U-EGDNet), comprising: The input module is used to acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map.

[0021] In one implementation, the input module takes a 640×640×3 RGB image of an underwater concrete structure as input; a DeXiNed pre-trained edge detection network is used to generate a 640×640×1 edge contour map that corresponds to the input RGB image at the pixel level, forming an RGB-Edge dual-stream input pair.

[0022] Among them, the DeXiNed pre-trained edge detection network is an advanced deep learning model designed specifically for edge detection. It can generate more refined and sharper edge maps, while also having good generalization ability.

[0023] The dual-stream feature extraction module is used to extract features from the RGB image and the edge contour map through the RGB semantic branch and the edge geometry branch respectively, and output the RGB multi-scale feature map and the edge multi-scale feature map.

[0024] Specifically, the dual-stream feature extraction module, as the core feature extraction unit of the dual-stream detection architecture of this invention, plays a key role in realizing multi-dimensional feature separation and extraction and hierarchical alignment of RGB images and edge contour maps, providing high-quality and highly adaptable feature inputs for subsequent hierarchical differentiated fusion.

[0025] The RGB semantic branch focuses on extracting the appearance semantic features, texture features, and global morphological features of defects from underwater concrete RGB images, capturing the semantic differences between defects and the background. The edge geometry branch focuses on extracting the boundary contour features, geometric morphological features, and weak edge detail features of defects from the edge contour map, enhancing the recognizability of defect boundaries. The two branches work in parallel and independently without interfering with each other, ensuring the purity of the two types of features. At the same time, it strictly ensures that the output RGB multi-scale feature map corresponds one-to-one with the edge multi-scale feature layer and matches the resolution. This provides the necessary prerequisite for the layered dual-stream fusion module to execute differentiated fusion strategies for different levels of features, effectively solving the technical pain point that single feature extraction in underwater scenes cannot take into account both defect semantics and boundary details.

[0026] It should be noted that, Figure 1 In the middle, Conv represents a convolutional layer.

[0027] In one implementation, both the RGB semantic branch and the edge geometry branch employ a lightweight convolutional neural network architecture, each containing four consecutive cascaded feature extraction levels. Each feature extraction level in the RGB semantic branch includes a cascaded structure of a CMDS module and a CTAM module. The input data to each feature extraction level is first processed by the CMDS module for feature extraction and then enhanced by the CTAM module. As the feature extraction level increases, deeper features are extracted, resulting in an output RGB multi-scale feature map forming a multi-scale feature pyramid from low-level fine-grained features to high-level global features. Each feature extraction level in the edge geometry branch includes a cascaded structure of an EEDC module and an EAEM module. The input data for each feature extraction level is sequentially processed by the EEDC and EAEM modules to obtain the output. The feature maps output by the two branches are perfectly aligned in terms of the number of levels and resolution, ensuring accurate fusion of subsequent mid- and high-level features at the same level without additional feature alignment processing, simplifying the fusion process while improving fusion accuracy.

[0028] The process involves several layers. The first feature extraction layer yields shallow features for the corresponding branch, with a size of 160×160. The second feature extraction layer processes these shallow features to obtain low-level features, with a size of 80×80. Subsequent layers yield mid-level features (40×40) and high-level features (20×20). It's important to note that shallow features only contain basic pixel textures and lack semantic and geometric information related to defects. Therefore, they cannot support defect classification or assist in accurate localization, contributing nothing to the final detection result. Consequently, only the corresponding low-, mid-, and high-level features are output in the RGB semantic and edge geometry branches; shallow features are not included in the output.

[0029] The hierarchical dual-stream fusion module is used to perform differential fusion of features at different levels in the RGB multi-scale feature map and the edge multi-scale feature map. This includes: generating attention information based on the mid-level edge features and applying it to the mid-level RGB features in a residual modulation manner to achieve edge-guided enhancement, and obtaining a first fused feature through adaptive weighted fusion; refining the first fused feature to obtain a mid-level fused feature; and fusing the high-level RGB features and the high-level edge features based on bidirectional attention multi-stream information interaction to obtain a second fused feature. The second fused feature is then modulated based on the spatial weights generated by the second fused feature to achieve feature calibration, resulting in a high-level fused feature.

[0030] Specifically, the layered dual-stream fusion module is the core innovative unit of this invention to solve the problem of insufficient accuracy in underwater concrete defect detection. Its core design logic is based on the different functional attributes of the middle and high-level features in RGB multi-scale features and edge multi-scale features, as well as the pain points of underwater defect detection scenarios. It adopts a differentiated fusion strategy to break the limitation of homogeneous fusion of multi-scale features in existing technologies.

[0031] The mid-level features focus on carrying fine-grained boundary and local detail information of defects, while the edge mid-level features have a clearer defect outline prior. Through edge attention generation and residual modulation, the response of key defect areas in the RGB mid-level features is enhanced in a targeted manner, and water noise interference is suppressed to achieve edge-guided enhancement. Then, through adaptive weighted fusion and feature refinement, redundant information is further removed and feature purity is improved, finally obtaining mid-level fusion features that take into account both boundary details and local semantics, providing support for accurate defect localization.

[0032] High-level features focus on carrying global semantic and overall morphological information of defects. Through bidirectional attention interaction between RGB high-level features and edge high-level features, deep complementarity and global semantic alignment of the two types of features can be achieved, avoiding semantic confusion caused by single features. Then, feature calibration is completed through spatial weight modulation to strengthen the defect region and suppress background interference, resulting in high-level fusion features with high semantic recognition and strong anti-interference ability, which provides a guarantee for accurate defect classification. The two types of fusion features work together to achieve simultaneous improvement in defect localization and classification accuracy.

[0033] In one implementation, the fusion process of RGB mid-layer features and edge mid-layer features is as follows: First, attention weights are generated in the channel dimension of the edge mid-layer features to obtain attention information that can characterize the saliency of defect edges. This attention information is then incorporated into the RGB mid-layer features in a residual modulation manner to achieve targeted enhancement of defect edge regions and suppression of background regions in the RGB mid-layer features. Then, the modulated RGB mid-layer features and edge mid-layer features are weighted and fused using a set learnable weighting coefficient to obtain the first fused feature. The first fused feature is then processed using a feature refinement method to remove redundant noise and invalid information introduced during the fusion process, and finally, the mid-layer fused feature is output.

[0034] Among them, feature refinement uses 1×1 convolution to integrate feature channels, combined with batch normalization and activation function to perform redundancy removal, noise reduction and feature regularization on the fused features, and suppress stray responses caused by water background without losing the details of defect edges.

[0035] The fusion process of RGB high-level features and edge high-level features is as follows: the RGB high-level features are used as query vectors and the edge high-level features are used as key vectors to generate the first set of attention maps. Then, the RGB high-level features are used as key vectors and the edge high-level features are used as query vectors to generate the second set of attention maps. The first set of attention maps is multiplied element-wise with the RGB high-level features to obtain the RGB interaction features. The second set of attention maps is multiplied element-wise with the edge high-level features to obtain the edge interaction features.

[0036] The second fusion feature is obtained by weighted fusion of RGB interaction features and edge interaction features. Global spatial features are extracted from the second fusion feature to generate spatial weights that represent the global distribution of defects. The second fusion feature is then modulated pixel by pixel using these spatial weights to complete feature calibration and finally output the high-level fusion feature.

[0037] The above fusion process does not require the introduction of additional complex feature conversion modules, thus ensuring the fusion effect while controlling computational overhead and adapting to the engineering application requirements of underwater detection.

[0038] The cross-scale fusion module is used to fuse mid-level fusion features, high-level fusion features, and RGB low-level features to output multi-scale detection features.

[0039] Specifically, the core function of the cross-scale fusion module is to integrate fusion features and original features from different levels and representation dimensions, breaking the limitations of each level of features (middle, high, and low) and constructing a unified feature expression system adapted to multi-scale defects in underwater concrete. Among them, the RGB low-level features carry the finest-grained texture and edge details of the defects, which is the basis for accurate defect localization. The middle-level fusion features take into account the defect boundaries and local semantics, while the high-level fusion features have strong anti-interference capabilities and clear global semantics. The fusion of the three achieves a synergistic effect of detail supplementation, boundary enhancement, and semantic support, effectively solving the technical problem that single-level features are difficult to adapt to defects of different scales such as underwater micro-cracks, local damage, and large-area spalling. At the same time, it removes redundant environmental noise in the features of each level, improves the representation capability of multi-scale detection features, and provides comprehensive and reliable feature input for the detection head module to accurately complete defect classification and localization.

[0040] In one implementation, the cross-scale fusion module is based on a feature pyramid fusion architecture. The specific process is as follows: High-level fusion features are upsampled using bilinear interpolation to adjust their resolution to match that of mid-level features (40×40). Then, they are concatenated with the mid-level fusion features. The concatenated features are then processed by the C3K2 module to obtain the mid-to-high-level fusion features. After upsampling, the mid-to-high-level fusion features are concatenated with the RGB low-level features. Finally, the C3K2 module outputs a first detection feature with the same size as the low-level features, balancing detail and semantics. The size is 80×80. The first detection feature is convolved with the mid-to-high-level fusion feature through a convolutional layer and then concatenated with the high-level fusion feature. The result is then concatenated with the high-level fusion feature through the C3K2 module to obtain the third detection feature, which has a size of 20×20. The entire process uses lightweight operation and does not require complex feature transformation. While ensuring fusion accuracy, it effectively controls the computational overhead and adapts to the actual engineering deployment needs of edge computing devices such as underwater robots.

[0041] The C3K2 module is a lightweight feature fusion module in YOLO11 based on an improved Cross-Stage Local (CSP) structure, representing an optimized upgrade of the traditional C2f module. This module employs a dual-branch design, integrating configurable multi-scale convolutional kernels into the processing branch for feature extraction. Through channel concatenation and feature regularization, it achieves efficient fusion and enhancement of multi-scale features, balancing feature representation capability and computational efficiency. It is commonly used in the neckline cross-scale fusion stage of object detection networks.

[0042] The detection head module is used to output the defect category, confidence level, and bounding box coordinates of the defect based on multi-scale detection features, through classification and regression branches.

[0043] In one implementation, the detection head structure in the YOLO11 architecture is adopted. For the multi-scale detection features output by the cross-scale fusion module, including the first detection feature, the second detection feature, and the third detection feature, corresponding to different resolutions, they are adapted to underwater concrete defects of different scales, such as micro-cracks, local damage, and large-area spalling. Independent classification and regression branches are configured for each size feature to achieve decoupling of defect classification and bounding box regression tasks and avoid feature optimization conflicts between the two tasks. Finally, the output results of the classification and regression branches of each scale are integrated to complete the end-to-end intelligent detection of underwater concrete defects and output the defect category, confidence level, and bounding box coordinates of the defect.

[0044] In summary, this invention employs a dual-stream feature extraction structure that combines RGB semantic branches and edge geometric branches, along with a differentiated hierarchical dual-stream fusion strategy designed for mid- and high-level features. Combined with cross-scale feature fusion and a decoupled detection head, it can effectively preserve fine-grained edge information of concrete defects and enhance the semantic discrimination ability of defects in underwater turbid and low-contrast environments. It is also adaptable to the detection needs of defects at multiple scales, from micro-cracks to large-area spalling. Overall, it solves the problem of insufficient detection accuracy caused by single feature fusion, easy loss of edges, and easy semantic confusion in existing technologies. While improving the reliability of defect localization and classification, it also takes into account the lightweight nature of the model and the practicality of engineering deployment.

[0045] To further illustrate the technical details of this invention, the technical solution will be described in detail below.

[0046] Both the RGB semantic branch and the edge geometry branch contain multiple levels of feature extraction structures, and each level outputs a feature map of the corresponding scale. In the RGB semantic branch, each level of feature extraction structure includes a CDMS module for multi-scale feature extraction and a CTAM module for channel response adjustment. In the edge geometry branch, each level of the feature extraction structure includes an EEDC module for preserving edge information and an EAEM module for spatial feature enhancement.

[0047] Specifically, the RGB semantic branch is constructed through concatenated convolutional layers, multiple CDMS modules, and multiple CTAM modules. The CDMS module is the Concrete Defect Multi-Scale Convolution Module, which includes a parallel four-branch structure: micro-scale branch, local branch, meso-scale branch, and macro-scale branch. The outputs of the four branches are concatenated along the channel dimension and then batch normalized and activated to obtain RGB multi-scale features. The four branches use convolutional kernels of different sizes to extract defect features under different receptive fields.

[0048] In one implementation, such as Figure 2 As shown, the CDMS module has four branches: micro-scale branch, local branch, meso-scale branch, and macro-scale branch. The micro-scale branch uses a 1×1 convolutional kernel to extract pixel-level texture features and outputs... The local branch uses a 3×3 convolution kernel to extract local neighborhood defect features and outputs... The mesoscale branch uses a 5×5 convolution kernel to aggregate mesoscale region features and output... The macro-scale branch employs asymmetric dilated convolution with kernels of 2×1 and 1×2 and a dilation rate of 2 to expand the receptive field and enhance the modeling capability of slender cracks. The output... .

[0049] After concatenating the outputs of the four branches along the channel dimension, and then applying batch normalized normalization (BN) and SiLU activation functions, multi-scale fused features are output. Specifically, it means: .

[0050] It should be noted that, Figure 2 In this context, SiLU represents the SiLU activation function layer.

[0051] The RGB semantic branch also includes the Channel Texture Attention Module (CTAM), whose structure is as follows: Figure 3 As shown.

[0052] In this implementation, the CTAM module divides the input features into two parallel processing paths. The main processing path sequentially passes through a 1×1 convolution and two 3×3 convolutional layers to model deep texture features and introduces channel attention (ECA) to adaptively recalibrate the channel responses. The other parallel path performs feature transformation through a 1×1 convolution to supplement lightweight channel information and enhance feature diversity. Finally, the two feature channels are concatenated and then output after adjusting the number of channels by a 1×1 convolution, expanding the number of channels to twice the number of input feature channels.

[0053] The CTAM module adaptively adjusts the channel response of input features through parallel feature transformation paths and channel attention mechanisms, suppressing the abnormal amplification of defect-irrelevant features and enhancing the response intensity of defect-related channels.

[0054] Figure 3 In the middle, GAP is a global average pooling layer, and Sigmoid is a Sigmoid activation function layer.

[0055] The edge geometry branch includes multiple EEDC modules, EAEM modules, SPPF modules and C2PSA modules connected in series, where the EEDC module includes a feature extraction branch and an edge preservation branch; The feature extraction branch uses convolutional layers to downsample the input features to obtain basic semantic features; The edge-preserving branch processes the input features sequentially through a feature transformation layer and a depthwise separable convolutional layer to preserve edge response information during downsampling.

[0056] like Figure 4 As shown, the EEDC module also adopts a dual-branch structure. The feature extraction branch uses a 3×3 convolution with a stride of 2 and padding of 1 to output standard features. The edge preservation branch EConv is constructed through 1×1 convolution, normalization, SiLU activation and depthwise separable convolutional blocks to reduce the smoothing damage of downsampling to weak edges and output edge-preserving features. The outputs of the two paths are concatenated along the channels and then subjected to batch normalization and SiLU activation function to output the final feature map.

[0057] The depthwise separable convolutional block consists of a cascaded depthwise convolutional layer (DWConv), a batch normalization layer, a pointwise convolutional layer (PWConv), and a batch normalization layer. Figure 4DWConv is a channel-wise convolution that independently extracts spatial features from each channel of the input feature map, retaining only the spatial texture and edge detail features of each channel without cross-channel information interaction; PWConv is a 1×1 pointwise convolution that performs cross-channel fusion on the multi-channel spatial features output by DWConv, and can adaptively adjust the number of output channels to achieve flexible control of feature dimensions.

[0058] By leveraging the collaborative efforts of DWConv and PWConv, the module significantly reduces the number of parameters and computational overhead while ensuring the ability to extract concrete defect features, thus adapting to the engineering deployment needs of edge computing devices such as underwater robots.

[0059] The SPPF (Spatial Pyramid Pooling Fast) module is a fast spatial pyramid pooling module used for efficient multi-scale feature aggregation. It significantly improves computational efficiency while maintaining a large receptive field, and is suitable for feature extraction needs of targets such as cracks and holes of different sizes in underwater concrete defect detection.

[0060] C2PSA (Convolutional Block with Parallel Spatial Attention) is a feature enhancement module that integrates the CSP (Cross Stage Partial) structure and the PSA (Partial Spatial Attention) mechanism. It dynamically focuses on key regions and suppresses background noise, making it particularly suitable for detecting small and blurry defects in complex underwater backgrounds. This module divides the input features into two paths: one path directly connects to retain the original information, and the other path enters the PSA attention branch. The two paths are fused through a 1×1 convolution, and finally, the enhanced feature map is output.

[0061] The edge geometry branch also includes the Edge-Aware Enhancement Module (EAEM); such as... Figure 5 As shown, the EAEM module incorporates a lightweight spatial attention mechanism (LSA). By calculating the mean and maximum values ​​of the input features along the channel dimension, it compresses the multi-channel features into a two-dimensional spatial response map. Then, it captures the spatial continuity of the defect edges through a 7×7 convolution, generates spatial attention weights, and applies them to the input features, explicitly enhancing significant edge regions and suppressing background noise responses.

[0062] In the hierarchical dual-stream fusion module, the processing of RGB mid-layer features and edge mid-layer features includes: Feature transformation is performed on the mid-layer edge features to generate an edge attention map. The mid-layer RGB features are then weighted based on the edge attention map. The weighted result is added element-wise to the mid-layer RGB features to obtain the edge guidance enhancement map. The edge guidance enhancement map and the mid-layer edge features are processed by depthwise separable convolutional blocks, and then added element-wise with the mid-layer RGB features and the mid-layer edge features respectively, and then adaptively weighted and fused to obtain the first fused feature. The first fusion feature is refined by a three-dimensional joint attention mechanism to generate a three-dimensional joint attention, including channel attention weights, height direction attention weights, and width direction attention weights. The three-dimensional joint attention is then weighted element-wise with the initial fusion feature to obtain the middle-layer fusion feature.

[0063] Among them, the first fusion feature The generation process is as follows: ; in, and These are RGB adaptive fusion weights and edge adaptive fusion weights, respectively. For depthwise separable convolutional blocks; Enhanced edge-guided graph; This is a mid-layer feature of RGB; This represents the features of the middle layer at the edge.

[0064] In one implementation, the fusion of RGB mid-layer features and edge mid-layer features includes edge-guided enhancement, adaptive weighted fusion, and feature refinement.

[0065] like Figure 6 As shown, the feature transformation in the edge-guided enhancement process involves transforming the mid-layer features of the edge. The edge attention map is obtained by sequentially performing 1×1 convolution dimensionality reduction, batch normalization, ReLU activation, and 1×1 convolution dimensionality increase, and finally Sigmoid activation. A learnable modulation coefficient was set. The edge attention map is linearly weighted, and the weighted result is compared with the mid-layer features of RGB. After element-wise multiplication, the result is then residually concatenated with the original RGB mid-layer features to obtain the edge-guided enhancement map. The calculation process is expressed as follows: , in, It is continuously updated during the training iteration process.

[0066] exist Figure 6In the adaptive weighted fusion process shown, the edge-guided enhancement map and the mid-layer edge features are extracted using depthwise separable convolutional blocks, and then residual connections are made with the mid-layer RGB features and the mid-layer edge features, respectively. The results of the two residual connections are then processed through learnable parameters. and Linear weighting and element-wise addition are used to complete the adaptive weighted fusion of the two features, resulting in the first fused feature. .

[0067] like Figure 7 As shown, feature refinement refines the first fused feature through a three-dimensional joint attention mechanism of channels, height, and width, generating channel attention weights respectively. Height-oriented attention weights Attention weights in width direction The first fusion feature is multiplied element by element to obtain the middle-layer fusion feature.

[0068] Figure 7 In this context, “Avgpool_w”, “Avgpool_h”, and “Avgpool” represent the average pooling layer along the width direction, the average pooling layer along the height direction, and the global average pooling layer, respectively.

[0069] In the hierarchical dual-stream fusion module, the processing of RGB high-level features and edge high-level features includes: Query vectors, key vectors, and value vectors are constructed based on RGB high-level features and edge high-level features. RGB interaction features and edge interaction features are obtained through multi-stream information interaction based on bidirectional attention. The RGB interaction features and the edge interaction features are weighted and fused to obtain the second fused feature; An edge saliency map is generated based on the second fusion feature. The edge saliency map is then concatenated with the second fusion feature channels. The channel concatenation result is processed to obtain spatial feature calibration weights. These spatial feature calibration weights are then used to calibrate the second fusion feature to obtain the high-level fusion feature.

[0070] The feature calibration includes at least two parallel calibration branches constructed based on the second fused feature: The first calibration branch is used to fuse the edge saliency map and spatial feature calibration weights, and to enhance the second fused feature to obtain the first feature calibration map. The second calibration branch is used to perform weighted modulation of the second fused feature based on the spatial feature calibration weight to obtain the second feature calibration map; The first and second calibration maps are fused together, and high-level fusion features are generated based on residual connections.

[0071] The implementation of feature calibration includes: The first calibration branch is constructed based on the edge saliency map, spatial feature calibration weights, and the second fusion feature: , in, This is the first feature calibration map. This is the second fusion feature. The edge enhancement coefficient, This is a saliency map of the edges. Adjust the weights for spatial features; A second calibration branch is constructed based on spatial feature calibration weights and second fusion features: , For the second feature calibration map; The first and second feature calibration maps are fused and then processed by convolution to obtain the high-level fused features. : , in, It is a 1×1 convolution.

[0072] In one implementation, the processing of high-level features includes multi-stream information interaction and feature calibration, wherein the multi-stream information interaction includes, for example, multi-stream information interaction. Figure 8 As shown, it specifically includes: Extracting high-level features of RGB respectively Features of the edge and high level On the one hand, using RGB high-level features as query vectors and edge high-level features as key vectors, edge information is injected into RGB high-level features through an attention mechanism to obtain RGB interactive features; on the other hand, using RGB high-level features as key vectors and edge high-level features as query vectors, semantic information is injected into edge high-level features to obtain edge interactive features.

[0073] The RGB interaction features and edge interaction features are dimensionally reshaped and subjected to global average pooling (GAP) respectively, and then concatenated along the channel dimension. The concatenated results are then normalized by a two-layer perceptron and Softmax to generate RGB adaptive fusion weights and edge adaptive fusion weights. and The two adaptive fusion weights are element-wise multiplied with the RGB interaction features and edge interaction features obtained from global average pooling, respectively, and then concatenated along the channel dimension. Finally, the second fusion feature is output through Fusion Block. The Fusion Block consists of convolutional layers, BN batch normalization layers, and SiLU activation layers.

[0074] like Figure 9As shown, during the feature calibration stage, the edge saliency map is extracted from the second fusion feature. The edge saliency map enhances the display of edges based on the second fusion feature. It concatenates the edge saliency map with the second fusion feature channels, and then processes the data through a convolutional layer to generate spatial feature calibration weights. Weights are calibrated based on spatial features, and weighted fusion is performed using a two-branch approach. Specifically: One branch involves edge enhancement coefficients and edge saliency maps. Second fusion feature The other branch is through a second fusion feature, which is weighted by spatial feature calibration weights. The weighted values ​​of the two branches are directly weighted with the spatial feature calibration weights. The weighted results of the two branches, namely the first feature calibration map and the second feature calibration map, are added element-wise. After adjusting the number of channels through a 1×1 convolution, the second fused feature is calibrated using a residual connection to obtain the high-level fused feature. The specific calculation process is as follows: , , , in, These are the edge enhancement coefficients, which are iteratively updated during model training.

[0075] This application also provides a method for detecting defects in underwater edge-guided dual-flow concrete, including: Acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map; Feature extraction is performed on the RGB image and the edge contour map by using the RGB semantic branch and the edge geometry branch respectively, and the output RGB multi-scale feature map and edge multi-scale feature map are generated. Differential fusion is performed on features at different levels in RGB multi-scale feature maps and edge multi-scale feature maps, including: generating attention information based on mid-level edge features and applying it to mid-level RGB features in a residual modulation manner, followed by adaptive weighted fusion; refining the fusion result to obtain mid-level fused features; and fusing high-level RGB features and high-level edge features after performing bidirectional attention interaction to obtain fused features, and modulating the fused features based on the spatial weights generated by the fused features to achieve feature calibration, thereby obtaining high-level fused features. The mid-level fusion features, high-level fusion features, and RGB low-level features are fused to output multi-scale detection features; Based on multi-scale detection features, the system outputs the defect category, confidence level, and bounding box coordinates of the defect through classification and regression branches.

[0076] To further evaluate the comprehensive performance of the underwater edge-guided dual-flow concrete defect detection network (U-EGDNet), comparative experiments will be conducted. Using mAP50 as the evaluation metric, supplemented by the mAP50-95 metric, the comprehensive detection performance of the network model under different positioning accuracy constraints will be analyzed, fully demonstrating the model's detection stability and robustness under complex underwater imaging conditions. This comparative experiment covers various representative target detection methods, including the classic two-stage detection framework, mainstream single-stage lightweight models, and detection methods designed for multimodal feature fusion in recent years. Through multi-dimensional comparison, the comprehensive advantages of the proposed method in terms of detection accuracy, model complexity, and multi-flow information utilization are systematically verified. The comparative experiments were conducted under three settings: RGB single-flow, Edge single-flow, and RGB and Edge dual-flow fusion. The experimental results are as follows: Figure 10 As shown.

[0077] Under RGB single-stream settings, most detection models achieved relatively stable detection performance, with the YOLO series models outperforming traditional two-stage methods. The mAP50 of YOLO12n, Mamba-YOLO, and RT-DETR reached 72.1%, 72.2%, and 71.8%, respectively. However, their mAP50-95 was generally below 52%, indicating that model performance degrades to some extent under high-precision positioning requirements. Although RGB single-stream performed well on cracks and corrosion defects, detection accuracy was significantly limited for defects such as exposed ribs that rely on fine-grained structural information, demonstrating the limitations of single RGB features under underwater illumination attenuation and background interference conditions.

[0078] Under the Edge single-stream setting, the overall detection performance of the model decreased compared to the RGB single-stream setting. Specifically, the mAP50 of RT-DETR was 59.3%, while the mAP50 of other YOLO series models mainly ranged from 56% to 63%, with corresponding mAP50-95 values ​​generally below 46%. From a category perspective, the Edge single-stream still maintained high detection accuracy for crack categories, validating the advantages of Edge features in characterizing slender cracks and structural boundaries; however, due to the lack of texture and semantic information, its ability to discriminate regional defects such as corrosion and spalling was significantly insufficient.

[0079] Under the RGB and Edge multi-stream fusion settings, the mAP50 of most fusion methods is improved compared to the single-stream settings, reflecting the complementarity between semantic information and geometric priors. DE-YOLO achieves an mAP50 of 74.9% and an mAP50-95 of 53.9%. Among all compared methods, the U-EGDNet proposed in this invention exhibits the best overall performance and efficiency, with mAP50 and mAP50-95 reaching 76.3% and 54.7% respectively, achieving best or near-best detection results across all four defect types. Furthermore, U-EGDNet has only 8.6M parameters and a computational cost of 20.2 GFLOPs, significantly lower than highly complex fusion models such as CFT and ICAFSion, demonstrating excellent accuracy-efficiency balance and engineering deployment potential.

[0080] Table 1 Ablation Experiment Results

[0081] To systematically evaluate the functional contributions of each key component in U-EGDNet, this invention conducted comprehensive ablation experiments. Considering that the framework proposed in this invention is a novel design, the baseline model is defined as a basic detection network without any enhancement modules or fusion mechanisms. This baseline model uses the YOLO11n backbone network and channel splicing as the feature fusion method. The experimental results are shown in Table 1. Without introducing any additional modules, the baseline model achieves mAP50 and mAP50-95 of 72% and 51.2%, respectively, which is better than the best results under the single-flow setting. Although the model has a low number of parameters and computational complexity, its ability to represent fine-grained structures and weak boundary targets in complex underwater environments is still limited.

[0082] In the single-branch ablation experiments, each key module demonstrated stable performance improvements. Introducing the RGB semantic backbone network increased the model's mAP50 to 73.5% and mAP50-95 to 52.9%, indicating that multi-scale semantic modeling and channel-texture-aware appearance feature extraction can effectively mitigate the adverse effects of texture degradation and scale changes in underwater environments. When only the Edge geometric backbone network was introduced, the model's mAP50 and mAP50-95 reached 73.1% and 52.5%, respectively. While slightly lower than the RGB semantic branch, these figures were still significantly better than the baseline model, validating the effectiveness of Edge geometric information in representing the boundaries of structural defects such as cracks and spalling.

[0083] In the dual-module combination experiment, the synergistic effect between different components was further verified. Directly combining the RGB semantic backbone network with the Edge geometric backbone network improved the model's mAP50 and mAP50-95 to 75.1% and 54.4%, respectively, representing improvements of 3.1% and 3.2% compared to the baseline model, fully demonstrating the complementary advantages between semantic appearance modeling and geometric boundary inference. When the RGB semantic backbone network and the hierarchical dual-stream fusion module worked together, the model's mAP50 and mAP50-95 reached 74.6% and 53.9%, respectively, achieving a relatively balanced performance between detection accuracy and computational efficiency. Similarly, the combination of the Edge geometric backbone network and the hierarchical dual-stream fusion module also achieved an mAP50 of 74.3% and an mAP50-95 of 53.4%, further demonstrating that the hierarchical dual-stream fusion module can effectively enhance the discriminative ability of single-stream features.

[0084] The complete U-EGDNet achieved state-of-the-art performance across all metrics, with mAP50 and mAP50-95 reaching 76.3% and 54.7%, respectively, representing improvements of 4.3% and 3.5% over the baseline model. Although the number of model parameters and computational complexity increased to 8.6M and 20.2 GFLOPs, respectively, this computational overhead remains highly feasible in practical underwater detection scenarios considering the significant improvement in detection accuracy.

[0085] The above embodiments are merely illustrative of the technical concept and features of the present invention, intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and should not be construed as limiting the scope of protection of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and thus all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

Claims

1. An underwater edge-guided dual-flow concrete defect detection system, characterized in that, include: The input module is used to acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map; The dual-stream feature extraction module is used to extract features from the RGB image and the edge contour map through the RGB semantic branch and the edge geometry branch respectively, and outputs the RGB multi-scale feature map and the edge multi-scale feature map. The hierarchical dual-stream fusion module is used to perform differential fusion of features at different levels in the RGB multi-scale feature map and the edge multi-scale feature map. This includes: generating attention information based on the mid-level edge features and applying it to the mid-level RGB features using residual modulation to achieve edge-guided enhancement; obtaining a first fusion feature through adaptive weighted fusion; refining the first fusion feature to obtain a mid-level fusion feature; and fusing the high-level RGB features and high-level edge features based on bidirectional attention multi-stream information interaction to obtain a second fusion feature; modulating the second fusion feature using spatial weights generated from the second fusion feature to achieve feature calibration, thus obtaining a high-level fusion feature. The cross-scale fusion module is used to fuse mid-level fusion features, high-level fusion features, and RGB low-level features to output multi-scale detection features; The detection head module is used to output the defect category, confidence level, and bounding box coordinates of the defect based on multi-scale detection features, through classification and regression branches.

2. The underwater edge-guided dual-flow concrete defect detection system according to claim 1, characterized in that, Both the RGB semantic branch and the edge geometry branch contain multiple levels of feature extraction structures, and each level outputs a feature map of the corresponding scale. In the RGB semantic branch, each level of the feature extraction structure is equipped with a CDMS module for multi-scale feature extraction and a CTAM module for channel response adjustment. In the edge geometry branch, each level of the feature extraction structure is equipped with an EEDC module for preserving edge information and an EAEM module for spatial feature enhancement.

3. The underwater edge-guided dual-flow concrete defect detection system according to claim 2, characterized in that, The CDMS module includes a parallel four-branch structure, namely micro-scale branch, local branch, meso-scale branch and macro-scale branch. The outputs of the four branches are spliced ​​along the channel dimension and then batch normalized and activated to obtain RGB multi-scale features. The four branches use convolution kernels of different sizes to extract defect features under different receptive fields.

4. The underwater edge-guided dual-flow concrete defect detection system according to claim 2, characterized in that, The EEDC module includes a feature extraction branch and an edge preservation branch; The feature extraction branch uses convolutional layers to downsample the input features to obtain basic semantic features; The edge-preserving branch processes the input features sequentially through a feature transformation layer and a depthwise separable convolutional layer to preserve edge response information during downsampling.

5. The underwater edge-guided dual-flow concrete defect detection system according to claim 1, characterized in that, In the layered dual-stream fusion module, the processing of RGB mid-layer features and edge mid-layer features includes: The edge mid-layer features are transformed to generate an edge attention map. The RGB mid-layer features are weighted based on the edge attention map. The weighted result is added to the RGB mid-layer features element by element to obtain the edge guidance enhancement map. The edge guidance enhancement map and the mid-layer edge features are processed by depthwise separable convolutional blocks, and then added element-wise with the mid-layer RGB features and the mid-layer edge features respectively, and then adaptively weighted and fused to obtain the first fused feature. The first fusion feature is refined using a three-dimensional joint attention mechanism to generate a three-dimensional joint attention, including channel attention weights, height direction attention weights, and width direction attention weights. The three-dimensional joint attention is then weighted element-wise with the initial fusion feature to obtain the intermediate fusion feature.

6. The underwater edge-guided dual-flow concrete defect detection system according to claim 5, characterized in that, The first fusion feature The generation process is as follows: ； in, and These are RGB adaptive fusion weights and edge adaptive fusion weights, respectively. For depthwise separable convolutional blocks; Enhanced edge-guided graph; This is a mid-layer feature of RGB; This represents the features of the middle layer at the edge.

7. The underwater edge-guided dual-flow concrete defect detection system according to claim 1, characterized in that, In the layered dual-stream fusion module, the processing of RGB high-level features and edge high-level features includes: Query vectors, key vectors, and value vectors are constructed based on RGB high-level features and edge high-level features. RGB interaction features and edge interaction features are obtained through multi-stream information interaction based on bidirectional attention. The RGB interaction features and the edge interaction features are weighted and fused to obtain the second fused feature; An edge saliency map is generated based on the second fusion feature. The edge saliency map is then concatenated with the second fusion feature channel. The channel concatenation result is processed to obtain a spatial feature calibration weight. The spatial feature calibration weight is used to calibrate the second fusion feature to obtain a high-level fusion feature.

8. The underwater edge-guided dual-flow concrete defect detection system according to claim 7, characterized in that, The feature calibration includes at least two parallel calibration branches constructed based on the second fusion feature: The first calibration branch is used to fuse the edge saliency map and spatial feature calibration weights, and to enhance the second fused feature to obtain the first feature calibration map. The second calibration branch is used to perform weighted modulation of the second fused feature based on the spatial feature calibration weight to obtain the second feature calibration map; The first and second calibration maps are fused together, and high-level fusion features are generated based on residual connections.

9. The underwater edge-guided dual-flow concrete defect detection system according to claim 8, characterized in that, The implementation of the feature calibration includes: The first calibration branch is constructed based on the edge saliency map, spatial feature calibration weights, and the second fusion feature: ， in, This is the calibration map of the first feature. This is the second fusion feature. The edge enhancement coefficient, For the edge saliency map, The weights are calibrated for the spatial features; A second calibration branch is constructed based on spatial feature calibration weights and second fusion features: ， This is the calibration map for the second feature; The first and second feature calibration maps are fused and then processed by convolution to obtain high-level fused features. : ， in, It is a 1×1 convolution.

10. A method for detecting defects in underwater edge-guided dual-flow concrete, based on the underwater edge-guided dual-flow concrete defect detection system according to any one of claims 1-9, characterized in that, include: Acquire the RGB image of the underwater concrete to be detected and generate the corresponding edge contour map; Feature extraction is performed on the RGB image and the edge contour map by using the RGB semantic branch and the edge geometry branch respectively, and the output RGB multi-scale feature map and edge multi-scale feature map are generated. Differential fusion is performed on features at different levels in the RGB multi-scale feature map and the edge multi-scale feature map, including: generating attention information based on the mid-level edge features and applying it to the mid-level RGB features in a residual modulation manner to achieve edge guidance enhancement, and obtaining a first fused feature through adaptive weighted fusion; refining the first fused feature to obtain a mid-level fused feature; fusing the high-level RGB features and the high-level edge features based on bidirectional attention multi-stream information interaction to obtain a second fused feature; modulating the second fused feature based on the spatial weights generated by the second fused feature to achieve feature calibration, and obtaining a high-level fused feature; The mid-level fusion features, high-level fusion features, and RGB low-level features are fused to output multi-scale detection features; Based on multi-scale detection features, the system outputs the defect category, confidence level, and bounding box coordinates of the defect through classification and regression branches.

Images