Weakly-Supervised Defect Segmentation Within Visual Inspection Images of Liquid Crystal Displays in Array Process

Abstract

This paper proposes a novel weakly-supervised defect segmentation approach for the visual inspection of LCD array images by a joint application of known weakly-supervised segmentation and unsupervised anomaly segmentation technique. Potential defect regions are firstly identified via active heatmap and masked out. Generative CAE trained via a GAN framework is further applied to create defect-free contents within the masked regions. A comparison between the generated image and the original image leads to precise defect segmentation. Our experiment over LCD inspection images showed that the proposed approach achieved comparable segmentation performance to the fully supervised FCN model, justifying its applicability for serious industry scenarios. Although the approach was motivated by and validated for the visual inspection in the LCD panel industry, it is conceptually a general method and has the potential to be applied in wide areas.