CNN-based network with multi-scale context feature and attention mechanism for automatic pavement crack segmentation

The diversity and complexity of cracks pose significant challenges for the rapid and accurate detection of pavement defects. To address these challenges, this paper aims to enhance feature utilization and develop an end-to-end crack segmentation network (CSNet), with the goal of significantly improving detection accuracy. Firstly, the proposed model integrates dense parallel dilated convolutions, enabling it to capture local information across multiple scales effectively. Secondly, an innovative multiscale context fusion module, combined with an attention mechanism, is introduced to effectively aggregate deep features, enhancing the perception of cracks. Finally, a generalized dice loss function is employed to further improve the training efficiency. Extensive experiments were conducted on three public datasets, and a comprehensive comparison was made with mainstream segmentation models. The results demonstrate that the proposed CSNet performs outstandingly across multiple evaluation metrics, achieving the highest F1-score of 0.7968 and mIoU of 0.8094, significantly surpassing other advanced segmentation models.