BRIDGE UNDERWATER STRUCTURAL CRACK DETECTION BASED ON IMAGE ENHANCEMENT AND IMPROVED U-NET FUSION

To address the challenges imposed by the detection of underwater structural cracks in bridges, such as manual detection difficulties, low precision, scarcity and low quality of crack images, and the dearth of effective features, this paper proposes a bridge underwater structural crack detection method based on image enhancement and improved U-net fusion. Low-quality underwater structural crack images collected in turbid water environments are augmented and then input into the UDnet network for enhancement, addressing problems such as low image clarity and poor contrast. A Convolutional Block Attention Module (CBAM) is incorporated into the U-net network to enhance the network's feature extraction capability. To tackle the issue of imbalanced positive and negative samples in the dataset, the loss function is modified to a combination of Focal Loss and Dice Loss, improving the network's learning effectiveness and generalization. Experimental results demonstrate that the improved network, with higher-quality underwater structural crack images input at the network's input end, enhances the network's detection and segmentation performance.