Improvements and experiments on defect detection of optical images of underwater structures

As the main form of hydraulic buildings, concrete structures often exhibit defects such as cracks, spallation, and exposed bars during their long-term service. This paper proposes an improved image enhancement algorithm for defect detection based on optical images, which integrates the dark channel map with the depth map and employs dynamic stretching of RGB channels and color compensation for color correction. Subsequently, a multi-source data set is constructed based on the yolov7 (You Only Look Once version 7) model for network training, and the Focal-EIoU (Focal-Efficient Intersection over Union) bounding box loss function is introduced to complete the object detection of defect image. Experimental results show that the detection accuracy of the improved algorithm has increased by $1.82 \%$ and $4.43 \%$, respectively, compared to YOLOv7 and YOLOv8 algorithms. Finally, the SGBM(Semi-Global Block Matching) algorithm is used to verify the accuracy of object positioning, and the results show that the relative error of object ranging is less than $6 \%$. To verify the feasibility of the improved algorithm, an optics-based defect simulation experiment was carried out in the laboratory pool environment. With multiple concrete test blocks as experimental objects, the comprehensive detection of underwater concrete defects is completed by combining the image enhancement algorithm, the improved yolov7 algorithm, and the ranging algorithm.