Global and Local Contrastive Learning for Classification and Segmentation of Mixed-type Wafer Bin Map Defect Patterns

Defect pattern recognition in semiconductor wafer bin maps (WBMs) presents a formidable challenge in the integrated circuit manufacturing industry. Precise wafer defect pattern classification and segmentation can trace the root cause of defect patterns in the manufacturing process, thereby mitigating cost losses and augmenting efficiency and quality of products. When different defects are mixed on the same wafer, the WBM becomes increasingly intricate, which further increases the difficulty of recognition. Existing supervised learning methods require a large number of labeled samples, which is undoubtedly labor-intensive. In this paper, we propose a self-supervised contrastive learning framework to classify and segment different mixed-type WBM defect patterns by combining global and local contrastive learning modules. Specifically, global contrastive learning module is designed to learn image-level representation, while local contrastive learning module is used to better understand the structure of local regions, which contributes to image segmentation task. Experimental results show that our model works well only requiring little labeled samples and abundant unlabeled samples.