Feature-Scale Attentive Pseudo-Labeling for Semi-Supervised Wafer Map Defect Segmentation

Accurate segmentation of wafer map defects is essential for ensuring the quality and reliability of semiconductor manufacturing. However, it is challenged by the complex nature of defect patterns and the scarcity of labeled data. This paper introduces a novel semi-supervised method for wafer map defect segmentation that leverages feature-scale attentive pseudolabeling. Our approach incorporates a cross-attention mechanism across multiple feature scales to enhance the quality of pseudolabels, thereby improving segmentation accuracy. By refining features at different scales, our method minimizes misclassification and accurately identifies diverse defect patterns. Experiments on the MixedWM38 dataset show our method substantially surpasses state-of-the-art techniques, achieving approximately 12 % higher mIoU and 10 % higher F1-score. Additionally, we evaluate our method on the Pascal VOC 2012 dataset to validate its generalization capability, achieving improvements of 5.6 % in mIoU and 3.2 % in Dice score. These results highlight the potential of our approach to significantly improve the precision and reliability of wafer map defect segmentation in semiconductor fabrication.