An enhanced two-stage intelligent crack detection method for concrete structures using improved YOLOX and U-Net3+ models

Concrete surface cracks threaten the durability and safety of structures. To overcome the inefficiencies and limited accuracy of manual inspection and traditional image-processing techniques, a two-stage algorithm based on deeply optimized YOLOX and U-Net3+ models is proposed. The YOLOX stage integrates an Efficient Channel Attention (ECA) module to enhance sensitivity to crack features, enabling rapid recognition and localization. In the U-Net3+ stage, Coordinate Attention (CA) and SimAM non-parametric mechanisms are incorporated, max pooling is replaced with Softpooling, ReLU activation is upgraded to Mish, and multi-scale separable convolutions are embedded within an eight-neighborhood node mask for more precise pixel-level segmentation. A skeleton-curve-based quantification algorithm predicts crack length and width. After distortion correction and scale transformation, maximum width, length, and area errors are below 0.05 mm, 1.5 mm, and 1 mm², respectively, compared to actual measurements. Experiments demonstrate a 69.87 % reduction in processing time relative to single-stage methods, significantly enhancing detection efficiency and accuracy.