Efficient detection of highway pavement cracks using computer vision-based MSFNet model

Most highway pavement cracks are long and thin with rich edge details. To balance the needs of pavement crack detection accuracy and real-time performance, while also accounting for local and global feature recognition, a self-focus lightweight crack segmentation method MSFNet based on multilevel correction is proposed. It uses blueprint separable convolution(BSConv) and MobileViT block to build a lightweight local and global crack feature extraction network, and incorporates a Self-Focus module to avoid feature information loss; Furthermore, to automatically correct the restoration of feature maps at different levels, we design a multilevel correction decoder(MCD) based on the coordinate-aware fusion(CAF) module; Then, combine it with a linear feature refining head to make the network's receptive field more consistent with the crack distribution in real scenarios. In addition, a knowledge distillation strategy based on feature-based and response-based is introduced to improve the performance and stability of the model through feature transfer and soft label constraints. Experimental results show that the mIoU of MSFNet on the three datasets of AerialCrack, CFD, and DeepCrack reaches 81.01%, 76.03%, and 87.34% respectively, which exceeds the best results of the comparison models by 1.56%, 0.88%, and 1.92% respectively. The F1 Scores reaches 88.42%, 84.43%, and 92.83% respectively, also achieving the best results. At the same time, the inference speed reaches 47.8FPS, which meets the actual application needs of pavement crack detection.