An Improved Crack Detection Method Based on Pyramid Split Attention

A small leak will sink a great ship. The existence of various cracks poses a huge threat to the quality and safety of engineering construction. For manual inspection of cracks, various complex cracks have high false detections or missed detections. Therefore, we proposed a crack detection algorithm based on attention mechanism to automatically locate cracks to improve the safety of the engineering construction. By improving the CSP layer in the backbone of YOLOv5, it introduces an attention mechanism called pyramid split attention (PSA) to focus on the importance of different channels to make the network better learn the crack features. In addition, during the testing process, we use TTA to synthesize the network detection performances at different scales to get the final detection results more accurately. Finally, the F1 score of our network reached a high level on the RDDC2020 dataset.