An asymmetric teacher-student network based industrial vision model for abnormal grain detection of semiconductor cooling devices

Unsupervised detection of micro-defects in thermoelectric cooler (TEC) grains faces significant challenges due to subtle, low-contrast anomalies and the high cost of manual annotation. In this work, we propose ATS-Net, an asymmetric teacher-student network built upon a shared ResNet backbone, designed as a lightweight, deployment-ready model that generalizes effectively across datasets without per-dataset tuning. Our key contributions include exponential moving-average normalization to stabilize feature statistics, a single-layer Real-NVP coupling mechanism to amplify teacher-student discrepancies at anomaly regions, and a dual-scale contextual transformer block facilitating joint local and global attention. ATS-Net is trained exclusively on defect-free samples in a two-stage process, and evaluated using image-level AUROC, pixel-level PRO at 95 % recall, and mean intersection-over-union (mIoU) metrics on the proprietary TEC-Grain dataset and the public MVTec-AD benchmark. Experimental results demonstrate that ATS-Net achieves superior performance, reaching 99.2 % AUROC, 99.1 % PRO, and 0.989 mIoU on TEC-Grain, and 98.8 % AUROC, 98.6 % PRO, and 0.97 mIoU on MVTec-AD. The model operates efficiently at 3.8 GFLOPs with 19.36 MB parameters and a speed of 96 FPS on an RTX 3090 GPU. Ablation studies show the introduced modules collectively enhance AUROC by 9.5 percentage points over the MKD baseline, with only a minor increase in parameters. ATS-Net thus effectively balances detection accuracy, interpretability, and speed, making it suitable for real-time defect inspection in semiconductor cooling device production. Future research will focus on integrating multi-modal fusion and self-supervised pretraining strategies to further minimize annotation requirements.