BRIDGE UNDERWATER STRUCTURAL DEFECTS DETECTION BASED ON FUSION IMAGE ENHANCEMENT AND IMPROVED YOLOV 7

This study presents a new method for automatic identification of structural defects by using ROV. First, low-quality fault photos are obtained under muddy water to expand the dataset. Next, the cascaded Water-Net underwater image improvement method is used to create high-quality images from low-quality photos. The polarized self-attention module is employed to preserve the high resolution output of the improved image, thus allowing the image enhancement and target detection to work together efficiently and enhance the detection accuracy, in order to counteract the hindrance caused by the discrepancy between image enhancement and target detection. Given that low-quality image data annotation would necessarily contain low-quality examples, the WioU loss function is used to replace the loss function in the original YOLOv7-tiny network model to improve generalization performance. The experimental results reveal that, compared with the original network, the revised network model retains defect recognition accuracy while retaining high picture contrast, sharper details, and superb visual effects, as well as dramatically improving missed detections and misjudgments.