Abstract: In order to improve the practicability of deep learning in X-ray weld defect detection and reduce the hardware requirements of defect detection task, the YOLO-M network with merely 3.6 M parameters is proposed. The amount of network calculation is reduced by introducing a lightweight inverse residual structure into the network; the multi-scale prediction mechanism is used to predict different defect characteristics by network layer; the positive sample information of the image is amplified across the grid to speed up the convergence speed in the process of network training. The network is not only applied to traditional computers, but also tested in the ultra-low power edge artificial intelligence chip Kanzhi K210. The results show that the detection accuracy of the proposed method is 93.5% and the detection speed is 11 fps. This method has good detection accuracy and greatly reduces the cost of defect quality inspection.