Bridge Crack Location Algorithm Based on Improved Fully Convolutional One-stage Detector

In order to solve the problem of slow locating speed in the process of bridge crack detection, a crack location algorithm based on fully convolutional one-stage(FCOS) detector is proposed. The algorithm uses the FCOS network model framework, uses the lightweight backbone network efficientnet to extract the characteristics of the crack image. And bi-directional feature pyramid network is introduced to fuse the characteristics of different scales of the crack image, thereby enhancing the visual feature extraction effect of the backbone network. Experiments on a self-made data set show that the algorithm can quickly locate bridge cracks. Compared with FCOS, the detection speed is increased by 10.4 frames per second while ensuring that the average accuracy is not attenuated. Compared with Faster-RCNN, Yolo4, Efficientdet, detection speed and average accuracy have obvious advantages.