🤖 AI Summary
This work addresses the challenge of missed detections, false positives, and misclassifications of subtle semiconductor lithography defects—such as bridging, burrs, necking, and contamination—in single-stage detection models. To overcome these limitations, the authors propose a two-stage vision-language detection framework. In the first stage, the Qwen3-VL model is efficiently fine-tuned using LoRA to perform defect counting, classification, and localization. The second stage introduces a failure-aware prediction refinement mechanism, which explicitly models and corrects initial prediction biases by training a refinement module on error cases from the first stage along with their corrected labels. This approach significantly reduces both false positive and miss rates while improving classification accuracy, outperforming existing single-stage fine-tuning strategies and achieving more robust and precise inference performance in lithography defect detection.
📝 Abstract
Semiconductor lithography inspection requires reliable detection of small pattern defects such as bridge, burr, pinch, and contamination. In this study, we propose a two-stage vision-language framework that combines initial defect detection with prediction refinement. In the first stage, Qwen3-VL is fine-tuned with LoRA as a vision-language adapter to predict defect counts, defect categories, and normalized bounding boxes from lithography images. However, direct fine-tuning may still produce common test-time errors, including false positives, missed defects, and incorrect defect types. To address this limitation, the second stage trains a refinement module using first-stage prediction failures and their corrected labels, allowing the model to review and revise initial outputs. By learning from cases where the initial adapter fails, the refinement process improves defect inference beyond single-stage fine-tuning.