A dual spatial and channel attention-based deep learning model for pavement crack segmentation

Road cracks must be promptly and accurately detected to protect public safety and maintain traffic infrastructure. Conventional image processing and typical convolutional neural network (CNN) techniques have limited efficacy in dynamic environmental conditions, particularly when segmenting faint, fragmented, and low-contrast fractures. This paper presents a unique lightweight dual attention network (DAN-Net) that incorporates both spatial and channel attention within a U-Net-based architecture, using residual blocks to address these problems. The proposed model dynamically highlights contextually meaningful spatial regions and feature channels, thereby enhancing segmentation accuracy across complex backdrops and variable lighting conditions. We assess DAN-Net on the publicly available CRACK500 dataset and demonstrate consistent improvements over established models, including U-Net, UNet++, ResUNet, SE-Net, and SCA-Net, across critical metrics. DAN-Net attains an F1 score of 70.97, a Jaccard index (JI) of 58.75, a precision of 88.76%, and a recall of 62.96%, surpassing all comparative approaches in F1, JI, and recall metrics. Additionally, it has a reduced parameter count (9.276 million), which enhances its suitability for real-time implementation. Experimental findings validate its resilience across various contexts and its capacity to markedly diminish false positives and manual review efforts.