When using yolov3 identification technology to detect the defects on the cable surface of cable-stayed bridge, although it has also achieved good detection results, there is a problem that the detection speed of the network is slow, and the detection accuracy of cracking is much lower than that of other kinds of defects. To solve these problems, this paper improves the running speed and detection accuracy, merges the BN layer and convolution layer, and embeds the CBAM attention mechanism into the yolov3 network. The experimental results show that the average accuracy map of crack detection is increased to 82.3%, and the running speed is increased to 51.4 frames per second, which meets the requirements of precision and speed of cable surface defect detection.