The detection of weld appearance defects of ship halls is mainly performed manually. Detection results depend on working experience, and working in ship halls is dangerous. In this paper, we design a crawling robot using the improved YOLOv5 to the detection of weld appearance defects on ship halls in real time. The features of weld appearance defects are not obvious and the detection objects are small. In addition, the attention mechanism is often used to enhance the features, and the small target detection layer can be used to strengthen small target recognition ability. Therefore, we improve YOLOv5 with CA and SPD-Conv to achieve good effect on detection of weld appearance defects. We apply the improved YOLOv5 model in the software of crawling robot, which makes the crawling robot detect the weld appearance defects on ship hulls in real time. The experimental results show that the improved model can effectively improve the mAP to increase by 0.026.