A lightweight weld defect recognition algorithm based on convolutional neural networks

This paper proposes a lightweight weld defect-recognition algorithm based on a convolutional neural network that is appropriate for weld defect recognition in industrial welding. Specifically, the developed scheme relies on the original SqueezeNet model. However, we improve the fire module to reduce the model’s parameter cardinality, introduce the ECA module to strengthen the learning of feature channels and improve the feature extraction ability of the overall model. The experimental results highlight that our algorithm’s average recognition rate on the overall defects of welding depressions, welding holes, and welding burrs reaches 97.50%. Note that although our model requires substantially fewer parameters, its recognition effect is significantly improved. Our algorithm’s feasibility is verified on the test data and challenged against current weld defect identification algorithms, demonstrating its enhanced identification role and application prospect.