Research on pavement crack detection based on improved YOLOv4

In the past, the rapid development of transportation infrastructure brought about a boom in transportation. Nowadays, a large number of transportation infrastructure has begun to enter the maintenance stage, especially a large number of highways have begun to need maintenance and repair in order to maintain normal operation and use. How to develop an efficient and reasonable maintenance plan needs to be judged according to the real road conditions on site. It is an important task to identify the existing damage on the road surface. At present, the detection of pavement cracks is inefficient and expensive. To solve this problem, an improved YOLOv4 pavement crack target detection model was proposed. Firstly, MobileNetv2 is used as the backbone network and other common convolution is replaced by deeply separable convolution. Secondly, coordinate attention mechanism and spatial attention mechanism are implanted into Backbone and Neck respectively. The experimental results show that the improved model can further improve the accuracy of pavement crack detection and greatly improve the detection speed, the FPS can reach 61.48 frames/s, and the mAP can reach 67.26%, which is greatly improved compared with the original model.