Pavement Crack Detection Based on Channel Attention Mechanism

Automated detection of pavement cracks plays a crucial role in road maintenance and traffic safety. One limitation of existing solutions is that they do not consider the different proportions of crack features and background noise in the feature channels extracted from the same pavement image by different convolution kernels. This paper improves the RT-DETR model achieving reducing a large number of parameters and better crack detection from complex background pavement images. Specifically, we first employ orthogonal channel attention to improve the backbone network ResNet50 so that it can enhance channels with high proportion of effective crack feature and suppress the channels with high noise proportion to batter extract multi-scale features. Second, we designed an efficient multi-scale feature fusion module with few parameters to fuse the small-scale features and the large-scale features through top-down and bottom-up pathway to obtain large-scale features and small-scale features, allow both of them can focus on crack areas. Finally, we conduct experiments on dataset China-D, and the experimental results show that our method can improve the FPS by about 31.5% and mAp50 by about 1.0%.