D2Seg: Dual-Image Differential Interleaved Semantic Segmentation Network for Accurate Semiconductor Wafer Surface Defect Detection

Abstract

Optical inspection of surface defects on patterned wafers plays a crucial role in yield management and control within advanced semiconductor manufacturing. Although several vision-based approaches have been proposed for this challenging task, they often struggle with fine edge segmentation, sensitivity to noise, and limited generalization to unseen defect types. To address these issues, we propose a novel dual-image-based segmentation network, termed D2Seg, for high-precision wafer surface defect inspection. The framework operates in two stages. First, a Backbone Reinforcement (BR) strategy is introduced to enhance the encoder’s robustness under low-contrast conditions, thereby improving the extraction of high-quality features. Second, a Siamese encoder-decoder architecture is designed to learn discriminative representations by comparing paired images, thereby enabling accurate defect-boundary segmentation and improved generalization. Specifically, we propose a Neighbor Feature Aggregation (NFA) module to achieve efficient and robust cross-scale and cross-location correlation modeling, along with high-level feature enhancement. Additionally, a set of Difference Information Extraction (DIE) modules is integrated to capture defect-related features at multiple scales, encouraging the network to suppress non-defect invariants while highlighting defect regions. Furthermore, a Scale Self-Attention Refinement (SAR) module is designed to augment the perception of defect areas, thereby refining the detection of subtle edges. Experimental results on two wafer-defect datasets collected from real production lines demonstrate that D2Seg outperforms the compared methods, achieving mIoU values of 94.12% and 87.97% and mPA values of 95.83% and 91.29%, respectively. These results demonstrate the robustness and effectiveness of the proposed method for detecting surface defects on patterned wafers, particularly in identifying small defects and producing precise edge segmentation.