Research on pavement crack locating based on style transfer and generative adversarial networks

This study presents a novel method for detecting pavement cracks using Class Activation Maps (CAMs). CAMs help identify specific regions in images that correspond to features like cracks in the pavement. The approach leverages style transfer learning with generative adversarial networks (GANs) to transfer the characteristics of crack images and normal images. This helps the model better understand and identify the features of pavement cracks. The generated CAMs, which highlight the locations of cracks, are used for accurate crack detection. To improve feature learning, the study uses adaptive global pooling, a technique that combines the benefits of both global average pooling and global max pooling. Additionally, the SE attention module is applied to refine the model's focus on key features by recalibrating the input data. This paper also proposes a method to enhance crack feature detection through style transfer. Experiments on four datasets—CFD, Crack500, Deepcrack, and WSIS—show that the proposed method significantly improves the quality of CAMs and helps the model more accurately detect pavement cracks. The results highlight the practical impact of this method, improving the efficiency of road maintenance by enabling more precise crack detection.