Bridge underwater crack detection method based on image enhancement

Abstract

The present application relates to the technical field of bridge structure safety detection and underwater image processing, solves the problems of underwater image degradation, low crack detection accuracy and difficulty in lightweight deployment of models. The present application adopts UWEMamba double-branch Mamba image enhancement network and UWCMamba lightweight segmentation network cascade, completes image enhancement through pixel-level and feature-level branch cooperation, and realizes crack pixel-level accurate segmentation relying on GSPM module, deformable convolution and coordinate attention. The present application has excellent detection accuracy, the model parameter quantity is only 3.94M, can be deployed on underwater robot edge equipment, is suitable for underwater structure crack intelligent detection of infrastructure such as bridge, port and underwater tunnel, replaces manual diving exploration, and improves detection efficiency and operation safety.

Description

I. Technical Field

[0001] This invention relates to the fields of bridge structural safety inspection, underwater image processing, and computer vision technology, specifically to a pixel-level detection method for underwater cracks in bridges based on dual-branch Mamba image enhancement and lightweight Mamba semantic segmentation. II. Background Technology

[0002] Underwater bridge structures, including core load-bearing components such as piers, abutments, and underwater piles, are constantly subjected to the impact of water flow, collisions with ships, chemical corrosion from water, and erosion by aquatic organisms during long-term service, making them highly susceptible to structural defects such as cracks. If underwater cracks are not detected and addressed in a timely manner, they will continue to expand, leading to a decrease in structural strength, and in severe cases, causing bridge collapse, posing a significant threat to transportation safety and public property.

[0003] Compared with land structure inspection, underwater crack detection of bridges faces two major technical bottlenecks that are difficult to overcome: First, underwater light scattering and absorption effects result in serious blue-green bias, low contrast, high noise from suspended particles, and blurred edges of fine cracks in the acquired images; second, underwater cracks are characterized by being long and continuous, having irregular geometric shapes, and having complex background interference, making it difficult for conventional detection methods to achieve accurate pixel-level segmentation.

[0004] Existing underwater image enhancement techniques have significant drawbacks: traditional non-physical model methods, such as histogram equalization and the Retinex algorithm, are prone to introducing noise and artifacts, and cannot preserve fine crack details; dark channel prior methods based on the Jaffe-McGlamery imaging model rely on prior parameter estimation and are prone to failure in turbid water; CNN-based deep learning models have limited receptive fields and are difficult to model global degradation features; Transformer-based models have high computational complexity and cannot meet real-time requirements; existing Mamba architecture enhancement models are general-purpose designs and are not optimized for underwater crack scenarios on bridges, failing to balance denoising and detail preservation.

[0005] Existing underwater crack detection technologies also have shortcomings: manual diving inspection is inefficient, risky, and highly subjective; target detection algorithms such as YOLO can only locate cracks and cannot provide pixel-level quantization data; CNN semantic segmentation networks have difficulty capturing the long-range continuity of cracks and are prone to missing cracks; Transformer segmentation models have a large number of parameters and cannot be deployed on edge devices such as underwater robots (ROVs); existing dedicated crack Mamba segmentation networks do not incorporate targeted image enhancement preprocessing, resulting in a significant drop in detection performance in low-quality underwater images.

[0006] In summary, the industry currently lacks a dedicated technical solution that can simultaneously address the three core issues of severe underwater image quality degradation, insufficient crack detection accuracy, and difficulty in deploying high-precision models in a lightweight manner, thus failing to meet the engineering requirements for intelligent and unmanned inspection of underwater bridge structures. III. Summary of the Invention

[0007] This invention addresses the shortcomings of existing technologies and aims to solve the following technical problems: resolving the technical contradiction of simultaneously achieving color shift correction, noise suppression, and preservation of fine crack details during underwater image enhancement; improving pixel-level segmentation accuracy of long, irregular, and low-contrast cracks in complex underwater environments, reducing false detection and false negative rates; and achieving a lightweight design of the detection model, enabling high-precision algorithms to run in real time on edge devices such as underwater robots.

[0008] An image enhancement-based method for detecting underwater cracks in bridges is proposed. The method adopts an end-to-end cascaded architecture and sequentially executes four steps: underwater image acquisition, UWEMamba dual-branch parallel image enhancement, UWCMamba lightweight crack segmentation, and outputting a binary crack segmentation mask.

[0009] The UWEMamba dual-branch parallel underwater image enhancement network first extracts shallow features from the original underwater bridge crack image through a 3×3 convolutional layer to generate an initial feature map, which is then simultaneously fed into the pixel-level branch PlbNet and the feature-level branch FlbNet for parallel processing. The feature-level branch FlbNet adopts an encoder-decoder architecture: the encoder consists of three cascaded SEMB-modified Mamba modules. Each SEMB module contains a spatially omnidirectional selective scanning SOSS unit and a multi-scale feedforward network MSFFN unit. The SOSS unit uses four-way omnidirectional scanning to enhance the continuity of the crack structure, and the MSFFN unit achieves multi-scale detail extraction through 1×1, 3×3, and 5×5 depth-separable convolutional branches. Each SEMB module is followed by a convolutional layer for downsampling. The decoder consists of a corresponding number of upsampling modules, which restore the resolution through deconvolution and fuse the corresponding stage features with the encoder through skip connections. The pixel-level branch of PlbNet maintains the original image resolution throughout, flattening the initial feature map into a one-dimensional pixel sequence. It models long-distance dependencies between pixels using the Mamba scanning mechanism, and then performs color correction and high-frequency detail enhancement via the channel-coordinate omnidirectional scanning CCOSS module, outputting pixel-level enhanced features. The dual-branch output features are concatenated along the channel dimension, weighted by residuals, and then passed through a 1×1 convolutional layer to output the enhanced RGB image.

[0010] The UWCMamba lightweight crack segmentation network enhances the image, which is then divided into fixed-size pixel blocks by the Patch Embedding module and converted into a one-dimensional feature sequence with position encoding. The encoder consists of four cascaded GSPM global scanning awareness Mamba blocks. Each GSPM block incorporates a two-dimensional selective scanning SS2D, a multi-directional serpentine scanning MSSS, a deformable convolutional DCN, and a pixel attention-guided fusion PAF unit. The MSSS uses a dual parallel serpentine and dual diagonal serpentine path scanning, and the deformable convolutional DCN adaptively matches the crack geometry. The decoder consists of four cascaded upsampling modules. After feature map upsampling, the modules are connected to the corresponding stage features of the encoder via skip connections. The feature input coordinate attention CA module is fused, and position-aware weights are generated through 1D average pooling in the horizontal and vertical directions to enhance crack features and suppress isolated noise. Finally, a 1×1 convolutional classifier outputs a pixel-level binary segmentation mask for the underwater crack of the bridge.

[0011] During network training, UWEMamba employs a combination of pixel loss, structural similarity loss, and perceptual loss for optimization; UWCMamba uses a combination of Focal loss and Dice loss to balance the distribution of crack and background samples. The overall model is built on the PyTorch framework, with only 3.94M parameters, and can be deployed on the edge devices of underwater ROVs for real-time detection.

[0012] The enhancement effect is optimal. The dual-branch collaborative architecture achieves the unity of global color correction and preservation of local fine crack details, completely resolving the inherent contradiction between underwater image denoising and detail loss.

[0013] The detection accuracy is leading. On the self-made UnderWater Crack dataset, the mIoU reaches 0.8780, the F1 score reaches 0.8692, the precision is 0.8777, and the recall is 0.8608. All indicators are significantly better than the existing mainstream methods.

[0014] It is quantizable and deployable, with a model parameter count of 3.94M, a file size of 37MB, and a computational load of 26.43G FLOPs. It achieves extreme lightweight design while ensuring high accuracy, and is suitable for real-time operation at the edge of underwater robots.

[0015] With outstanding engineering value, it can completely replace manual diving exploration operations, significantly reducing detection costs and operational risks. It is suitable for underwater structural health monitoring of infrastructure such as bridges, ports, and underwater tunnels, and has broad prospects for industrial application. IV. Description of the attached drawings

[0016] Figure 1. Overall framework diagram of the image enhancement-based underwater bridge crack detection method of the present invention; Figure 2. Structure diagram of the UWEMamba dual-branch underwater image enhancement network of the present invention; Figure 3. Structure diagram of the SEMB improved Mamba module of the present invention; Figure 4. Structure diagram of the UWCMamba underwater bridge crack segmentation network of the present invention; Figure 5. Structure diagram of the GSPM global scanning perception Mamba block of the present invention; Figure 6. Visualization diagram of the underwater bridge crack detection effect of the present invention. V. Detailed Implementation Methods

[0017] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0018] The Underwater Image Augmentation Dataset UW-Pix is ​​based on three publicly available underwater image datasets: UIEB, EUVP, and LSUI. Samples of non-bridge scenes, such as marine life and corals, are removed, and samples of underwater structures, such as bridge piers and piles, are added to create a dedicated dataset. This dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The Underwater Crack Dataset contains 1037 real-world engineering images, covering different water depths, turbidity levels, lighting conditions, and crack morphologies. It includes complex interference such as suspended particles and attached organisms. Pixel-level binary annotations were completed by professional bridge inspectors. The dataset is divided in a 7:2:1 ratio and augmented using random flipping, rotation, and brightness adjustments.

[0019] The model is implemented using the Python 3.8 programming language and the PyTorch 2.0 deep learning framework, supports CUDA GPU accelerated computing, and the hardware environment consists of an Intel Core i9 CPU and an NVIDIA RTX 4090 graphics card.

[0020] UWEMamba network training: Input the original underwater image, output the enhanced and clear image; a combined loss function is used to integrate pixel-level error, structural similarity, and perceptual feature differences, updating parameters through backpropagation, and stopping when the validation set loss converges, saving the optimal model weights. UWCMamba network training: Input the UWEMamba enhanced image and the corresponding crack annotation mask; a combined Focal loss and Dice loss is used to address the imbalanced sample problem, and maximizing the validation set mIoU is used as the convergence condition, saving the optimal segmentation model.

[0021] On the UnderWater Crack test set, compared with eight mainstream methods (SFIAN, CTCrackSeg, SimCrack, PlainMamba, MambaIR, DCUFormer, DCCM-Net, and DBCNet), the method of this invention achieved the best performance across all metrics: mIoU 0.8780, ODS 0.8615, OIS 0.8677, precision 0.8777, recall 0.8608, and F1 score 0.8692. The model has 3.94M parameters, a model size of 37MB, and a computational cost of 26.43G FLOPs, demonstrating significant lightweight advantages.

Claims

1. A method for detecting underwater cracks in bridges based on image enhancement, characterized in that, include: The original underwater image of the bridge is obtained, and the original image is enhanced by the UWEMamba dual-branch parallel underwater image enhancement network to obtain the enhanced image. The enhanced image is then segmented at the pixel level by the UWCMamba lightweight crack segmentation network, and the binarized segmentation mask of the underwater crack of the bridge is output. The UWEMamba dual-branch parallel underwater image enhancement network includes a pixel-level branch PlbNet and a feature-level branch FlbNet, which are processed synchronously. The pixel-level branch PlbNet performs feature extraction and color correction at the original image resolution, while the feature-level branch FlbNet uses an encoder-decoder architecture to complete multi-scale semantic feature enhancement. The output features of the two branches are fused to obtain the enhanced image. The UWCMamba lightweight crack segmentation network includes a Patch Embedding module, an encoder, a decoder, and a segmentation head connected in sequence. The encoder is composed of stacked GSPM global scanning perception Mamba blocks, and the decoder integrates a coordinate attention module to complete feature optimization during the skip connection stage.

2. The detection method according to claim 1, characterized in that, The UWEMamba dual-branch parallel underwater image enhancement network first extracts shallow features from the original image through a 3×3 convolutional layer to generate an initial feature map, which is then input into the pixel-level branch PlbNet and the feature-level branch FlbNet for parallel processing. The FlbNet encoder, a feature-level branch, contains three cascaded SEMB-modified Mamba modules. Each SEMB module is followed by a convolutional layer to perform downsampling. The decoder contains an upsampling module corresponding to the encoder. The upsampling module restores resolution through deconvolution and performs skip connections to fuse the corresponding stage features of the encoder. The SEMB-modified Mamba module incorporates a spatially omnidirectional selective scanning (SOSS) unit and a multi-scale feedforward network (MSFFN) unit. The SOSS unit uses a four-way omnidirectional scanning strategy to enhance the continuity of the crack structure, while the MSFFN unit extracts multi-scale detailed features through 1×1, 3×3, and 5×5 depth-separable convolutional branches.

3. The detection method according to claim 2, characterized in that, The pixel-level branch PlbNet flattens the initial feature map into a one-dimensional pixel sequence, models the long-distance dependency of pixels through the Mamba scanning mechanism, and completes color correction and high-frequency detail enhancement through the channel coordinate omnidirectional scanning CCOSS module, outputting pixel-level enhanced features; The UWEMamba dual-branch parallel underwater image enhancement network concatenates pixel-level enhancement features and semantic-level enhancement features along the channel dimension, and outputs the enhanced RGB image through a 1×1 convolutional layer after residual weighted fusion.

4. The detection method according to claim 1, characterized in that, The Patch Embedding module divides the enhanced image into fixed-size pixel blocks and converts them into one-dimensional feature sequences with positional encoding, which are then input into the encoder of the UWCMamba network. The encoder of the UWCMamba network contains four cascaded GSPM global scanning perceptual Mamba blocks. Each GSPM block is followed by a downsampling operation. The GSPM block incorporates a two-dimensional selective scanning SS2D unit, a multi-directional serpentine scanning MSSS unit, a deformable convolutional DCN unit, and a pixel attention-guided fusion PAF unit.

5. The detection method according to claim 4, characterized in that, The MSSS unit employs a scanning strategy of dual parallel serpentine paths and dual diagonal serpentine paths. The sampling points of the deformable convolutional DCN unit adaptively match the geometric deformation of the crack, and the PAF unit completes the weighted fusion of scanning features and initial features.

6. The detection method according to claim 1, characterized in that, The decoder of the UWCMamba network contains four cascaded upsampling modules. The upsampling modules double the size of the feature map and then connect it to the corresponding stage features of the encoder via jump connections. The fused features are input to the coordinate attention module to complete noise suppression and crack feature enhancement. The coordinate attention module performs parallel horizontal and vertical 1D average pooling on the fused features to generate position-aware features. After convolution and activation, an attention weight map is generated, which is then multiplied element-wise with the original fused features to complete feature enhancement.

7. The detection method according to claim 1, characterized in that, The segmentation head uses a 1×1 convolutional layer to perform channel mapping on the decoder output features, generating a pixel-level binary segmentation mask for underwater cracks in the bridge.

8. The detection method according to claim 1, characterized in that, The UWEMamba network is trained using a combined loss function, which optimizes parameters by integrating pixel error, structural similarity, and perceptual feature differences. The UWCMamba network is trained using a combination of Focal loss and Dice loss to balance the distribution of crack and background samples. The UWEMamba and UWCMamba networks are built on the PyTorch framework, with an overall model parameter size of no more than 4M, enabling real-time detection on the edge devices of underwater ROVs.

Images