YOLO-MSPA: a multiscale polarized attention network for high-precision pavement crack segmentation

This paper proposes YOLO-MSPA, a multi-scale polarized attention network for pavement crack segmentation addressing high leakage rates, insufficient positioning, and environmental adaptability challenges. This network integrates ASPP (Atrous Spatial Pyramid Pooling) module and PSA (Polarized Self-Attention) mechanism, where the ASPP module capturing multi-scale contexts to enhance crack continuity perception, and the PSA mechanism optimizing channel-spatial weights to suppress background interference. Experimental comparisons verifed the superior performance of the proposed network, achieving 79.45% mIoU and 80.01% Precision in complex and low-contrast scenarios. The ablation analysis also confirmed the synergistic effect of ASPP and PSA.