A Transformer-based Attention Network for Crack Defect Detection

Pavement cracks have the characteristics of complex topological structure, uneven intensity and complex background. Therefore, accurate pavement crack detection is still a challenging problem for road maintenance. Since the rapid development of deep learning, detection methods based on Convolutional Neural Network (CNN) have achieved excellent results in crack detection tasks. However, the CNN-based model does not fully grasp the global information. Global information is crucial for accurate detection results. Transformer is powerful at capturing global information. Therefore, this paper fuses the CNN and Transformer framework to present an accurate crack detection model. To enhance the representation ability of the Transformer for global features, based on the improved self-attention mechanism, an enhanced Transformer unit is proposed. Meanwhile, an axis orientation attention module (AOAM) is proposed to optimize the extracted features at the encoder side, which realizes attention along both height and width directions through the self-attention mechanism. Experiments show that proposed model achieves an excellent results on two publicly available crack datasets.