Intelligent Detection of Sealed Crack with 2D Asphalt Pavement Images

Abstract

Accurately identifying sealed cracks on asphalt pavement surfaces is of significant importance to pavement management. This paper proposes an efficient semantic segmentation model called Parallel-TDNet for pixel-level detection of pavement sealed cracks. The proposed Parallel-TDNet presents two major modifications of the DeepLabv3+ model. First, the self-attention mechanism is applied at the end of the downsampling process to capture long-range dependency and enhance utilization of global information relationships. Second, a concurrent squeeze and excitation block is added to the original decoder of the DeepLabv3+ model to capture the details of sealed cracks. Experimental results demonstrate that the proposed Parallel-TDNet model on 932 testing images achieves a mean F-measure of 84.83% and a mean intersection-over-union of 0.7366 respectively. Compared with several efficient semantic segmentation models, such as PSPNet, FCN, SegNet, U-net, DeepLabV3+, SegFormer, the Parallel-TDNet algorithm yields a noticeably higher detection accuracy.