MPSU-Net: Quantitative Interpretation Algorithm for Road Cracks Based on Multiscale Feature Fusion and Superimposed U-Net

Road cracks pose a persistent challenge in road maintenance, with timely detection and repair crucial for enhancing road safety. However, determining which cracks require repair can be difficult, necessitating a quantitative analysis approach. This paper proposes a deep learning-based method, Multiscale Feature Fusion and Superimposed U-Net (MPSU-Net), for precisely this purpose. The method employs MPSU segmentation, which quantifies crack interpretation by analyzing the black and white pixels in binary images. Within the segmentation algorithm, Attention Connection is introduced to fuse features across different layers, while the PSU Block amalgamates feature information within the same layer, incorporating ASPP, CBAM, and the superimposed U-Net. The superimposed U-Net is designed to enhance PSU's feature extraction capabilities. Furthermore, a new Conv block is introduced to bolster feature extraction by replacing all convolutions in the superimposed U-Net and decoder. To address imbalanced positive and negative crack samples, we adopt the Jaccard loss metric based on experimental results. We enhance dataset diversity by leveraging data augmentation and amalgamating data from multiple datasets, resulting in the comprehensive Crack500 and Crack datasets. Experimental findings showcase the significant efficacy of MPSU-Net in enhancing F1-score and Recall metrics. On the Crack500 dataset, MPSU-Net achieves F1-scores and Recall rates of 85.32% and 78.68%, respectively, representing notable improvements of 6.19% and 6.43% over the U-Net baseline performance. Similarly, on the Crack dataset, MPSU-Net attains F1-scores and Recall rates of 74.54% and 46.66%, marking enhancements of 8.25% and 8.64% over the U-Net baseline. These results underscore the high performance of MPSU-Net and its potential to aid in quantitative road crack analysis.