Multi-stage few-shot micro-defect detection of patterned OLED panel using defect inpainting and multi-scale Siamese neural network

Automatic micro-defect detection is crucial for promoting efficiency in the production lines of patterned OLED panels. Recently, deep learning algorithms have emerged as promising solutions for micro-defect detection. However, in real-world industrial scenarios, the scarcity of training data or annotations results in a drop in performance. A multi-stage few-shot micro-defect detection approach is proposed for patterned OLED panels to deal with this problem. Firstly, we introduce a converter from defective to defect-free images based on our redesigned Vector Quantized-Variational AutoEncoder (VQ-VAE), aiming to inpaint defects with normal textures. Next, we exploit a region-growing method with automatic seed points to obtain the defect’s segmentation and geometric parameters in each image block. Reliable seed points are provided by structural similarity index maps between defective sub-blocks and reconstructed reference. Finally, a multi-scale Siamese neural network is proposed to identify the category of extracted defects. With our proposed approach, detection and classification results of defects can be obtained successively. Our experimental results on samples at different array processes demonstrate the superb adaptability of VQ-VAE, with a defect detection rate ranging from 90.0% to 96.0%. Additionally, compared with existing classification models, our multi-scale Siamese neural network exhibits an impressive 98.6% classification accuracy for a long-tailed defect dataset without overfitting. In summary, the proposed approach shows great potential for practical micro-defect detection in industrial scenarios with limited training data.