NRAET: a noise reduction and attention enhanced transformer for pavement damage classification in complex scenarios

Recent work in computer vision and deep learning has improved pavement damage detection and classification, but many models still work poorly in complex pavement scenes where damage areas are small, defect shapes are irregular, features appear at different scales, and background noise affects recognition. To solve these problems, this study proposes a Noise Reduction and Attention Enhanced Transformer (NRAET) for pavement damage classification. The NRAET is based on a Swin Transformer and has two main parts. The first is the Spatial Noise Reduction (SNR) module, which removes unnecessary information and reduces background noise. The second part is the Multi-Attention Enhancement Block (MAEB), which combines window, multi-scale, and fine attention to help the model better detect irregular cracks at different scales and learn global features. We conducted several experiments on several public benchmark datasets with different pavement types and damage patterns. For the CQU-BPMDD dataset, NRAET performed better than the baseline model. Compared with the baseline Swin Transformer, NRAET achieved improvements of 1.64% in Accuracy, 1.20% in Precision, 1.47% in F1-score, and 3.23% in AUC on the CQU-BPMDD dataset. The results show that NRAET under different data distributions has a good effect, very strong robustness, and generalization ability.