CFDSLite-UNet: A lightweight algorithm for pavement raveling distress detection with efficiency and accuracy synergistic improvement

Raveling distress is a prevalent form of early-stage deterioration in asphalt pavements. It accelerates the progression of other distress mechanisms and significantly affects both traffic safety and ride quality. Accurate and reliable detection of raveling distress is therefore critical to the development of digital road infrastructure management and intelligent driving environment perception systems. However, existing detection devices and algorithms exhibit several inherent limitations, including ambiguous texture variations and high computational complexity. To address these challenges, a lightweight, improved CFDSLite-UNet model for raveling distress segmentation and grading is proposed, incorporating higher resolution pixel feature capabilities. In particular, adaptive histogram equalization and morphological closing operations are applied to improve image contrast and enhance distress strengthening during preprocessing and training. Depthwise separable convolution and bilinear interpolation are adopted to improve model lightweight and enhance the convergence and robustness of the momentum-adaptive learning rate algorithm. Experimental results demonstrate that the proposed algorithm outperforms other mainstream CNN- and Transformer-based approaches in both computational efficiency and accuracy. The mIOU and F1-score are increased to 0.96 and 0.98, respectively. Meanwhile, the number of model parameters was reduced by approximately 88.5% which was optimized to 3.56 M, and the convergence is improved by 50%. The study contributes to the advancement of intelligent, high-precision pavement distress detection systems and supports fast speed data acquisition and transmission technologies for road infrastructure monitoring.