Extraction of crack characteristics and local damage detection in complex hydraulic concrete structures

Crack detection in complex hydraulic structures is often challenged by background interference, which obscures crack contours and hampers accurate damage localisation. This paper proposes a novel method to address this issue. First, an image compression technique integrating YCbCr conversion, visually weighted wavelet transform, and embedded zero-tree coding is introduced to enhance crack saliency and reduce redundancy. Subsequently, crack edges are continuously and precisely extracted using an adaptive dual-threshold Canny operator. Finally, pixel-level damage detection is achieved with an encoder-decoder model based on SegNet, trained via the Nadam optimiser. Experimental results demonstrate the superiority of our approach, achieving a detection accuracy of 96.2%, a mean intersection over union of 95.6%, and an F1-score of 0.97, significantly outperforming existing methods. The proposed technique provides a reliable and robust solution for the intelligent diagnosis of hydraulic engineering structures.