A real-time detection approach for bridge cracks based on YOLOv4-FPM

In order to realize real-time detection for bridge cracks by unmanned aerial vehicle (UAV), a deep learning model named YOLOv4-FPM is proposed on the basis of the YOLOv4 model. In YOLOv4-FPM, focal loss is used to optimize the loss function, which improves the accuracy and overcomes the challenges of complex background. Pruning algorithm is used to simplify the network and accelerate the detection speed. The multi-scale dataset is used to expand the predictable range of YOLOv4-FPM and enhance its scale robustness. The experimental results show that the mean average precision (mAP) of YOLOv4-FPM is 0.976, which is 0.064 higher than YOLOv4. The size and parameters of the model are reduced to 18.2%, and the model processes in real-time (119FPS) images at 1000 × 1000 pixels, which is 20 times faster than in a recent work. Moreover, it can effectively detect cracks in images of different sizes. • YOLOv4-FPM can detect cracks across complex background using focal loss • Focal loss can significantly improve the detection accuracy of the model • Pruning greatly improves the detection speed to meet the real-time requirements • Multi-scale training is necessary to improve the scale robustness of the model