Causal deep learning for explainable vision-based quality inspection under visual interference

Vision-based quality inspection is a key step to ensure the quality control of complex industrial products. However, accurate defect recognition for complex products with information-rich, structure-irregular and significantly different patterns is still a tough problem, since it causes the strong visual interference. This paper proposes a causal deep learning method (CDLM) to tackle the explainable vision-based quality inspection under visual interference. First, a structural causal model for defect recognition of complex industrial products is constructed and a causal intervention strategy to overcome the background interference is generated. Second, a defect-guided recognition neural network (DGRNN) is constructed, which can realize accurate defect recognition under the training of CDLM via feature-wise causal intervention using two sub-networks with feature difference mechanism. Finally, the causality between defect features and defective product labels can guide the DGRNN to complete the accurate and explainable learning of defect in a causal direction of optimization. Quantitative experiments show that the proposed method achieves recognition accuracy of 94.09% and 93.95% on two fabric datasets respectively, which outperforms the cutting-edge inspection models. Besides, Grad-CAM visualization experiments show that the proposed method successfully captures the data causality and realizes the explainable defect recognition.