96‐4: Self‐Supervised Outpainting for Display Panel Defect Image Augmentation

Deep learning (DL) has gained increasing popularity over conventional computer vision methods in smart manufacturing for its potential in high performance. However, training good DL models needs a large set of training data with sufficient variations. In the case of defect detection and classification for display panels, it is desirable for the defect (NG) images to have variations in defect shape, size and location. In practice, NG images are often collected without these considerations, e.g., defects are often located at the center. To increase data variation of the dataset, a novel image outpainting method is proposed to support expanding an input NG image to larger size so that data augmentations like random scales are applicable. A series of special masks are designed to support self‐supervised training of the outpainting model using NG images exclusively. More importantly, the same model can be applied to remove defect from an input NG image and output a matching normal (OK) image. These augmented OK‐NG pairs can benefit other DL methods, e.g., a defect image generation AI model can be trained as an OK conditional model to support generating defect on an input OK image while preserving the background. It is shown with experiments that a defect generation model, which is trained with augmented images using outpainting, is able to generate samples with higher defect quality and better background fidelity comparing to previous methods.