Detection of concrete bridge surface damage using wavelet-based multiband channel attention mechanism

This study proposes a dual improvement strategy and constructs a dedicated dataset to address the challenges of insufficient feature extraction, information loss, and high model complexity in detecting concrete bridge surface damage. A wavelet-based multiband channel attention mechanism (WMCAM) is developed to establish channel-wise attention weights through multiscale analysis of feature responses across different frequency bands, significantly enhancing damage feature extraction in complex backgrounds. Furthermore, an innovative compound convolutional fusion module (C2f-BNS) is introduced, integrating channel shuffle and pointwise convolution to enhance cross-channel information exchange while reducing model parameters by 53.6%. To overcome the limitations of existing datasets, the Chongqing concrete bridge surface damage dataset (CCBSD) is constructed. This dataset comprises 7,243 high-resolution images with expert annotations for four typical defect categories: CorrosionStain, ExposedBars, Efflorescence and Spallation. Experimental results demonstrate that the improved model achieves 72.9% in mAP50 on the CCBSD dataset, with the WMCAM and C2f-BNS modules contributing 2.4% and 1.4% performance gains, respectively. The proposed method effectively balances detection accuracy and computational efficiency through a 53.6% parameter reduction, providing a novel technical pathway for intelligent bridge inspection. This work aligns with practical engineering requirements whilst advancing computer vision applications in infrastructure health monitoring, particularly through its frequency-aware attention mechanism and lightweight architecture design.