Abstract: In order to solve the impact of ambient light and realize automatic rust removal of spliced welds on large ships, a complex cross-welding joint type discrimination method based on the EfficientNet deep learning model was proposed. Based on the image data of actual ship decks under different lighting conditions, ResNet and EfficientNet series models were used to learn image features, and a recognition model with an F1 score close to 1 was obtained. The results show that even under strong or weak light conditions, the model can still effectively identify seven types of cross welds and their transition states: straight line, cross, T-shaped, left L, right L, left T and right T. However, after using Openvino to accelerate deployment, EfficientNet's accuracy and time consumption are both better than ResNet, and the calculation time as low as 6.34 ms, which well meets the needs of real-time identification of cross welds and lays a foundation for autonomous rust removal path of wall-climbing rust removal robots.