Abstract Welding is a critical process in numerous industrial production scenarios. To ensure safe production, defects that occur during welding must be timely and effectively identified. Deep learning has been widely used in welding defect identification due to its powerful feature extraction capabilities, but in environments with limited computing resources, large deep learning models are difficult to deploy. Thus, a lightweight model for welding defect recognition should be devised while maintaining high accuracy under limited computational resources. Accordingly, we propose a weld defect recognition model called LSRSNet and based on a lightweight architecture. Specifically, based on the SqueezeNet model, our model uses the lightweight linear deformable convolution to simplify the model and improve the recognition accuracy. Moreover, a squeeze-and-excitation attention mechanism is integrated into the Fire module of the original SqueezeNet to enhance the feature expression ability, and a residual structure is constructed to prevent degradation of the deep network and improve the learning ability of complex features. Experimental results show that compared with the original model, the accuracy of the improved LSRSNet on a weld defect dataset increases by 9.5%, with notably fewer parameters than the original model. We offer a novel algorithmic for welding defect recognition and believe that LSRSNet can be deployed for industrial testing.
Li et al. (Tue,) studied this question.