Does Super-Resolution Preserve Defect Evidence? A Low-False-Call Benchmark for Semiconductor Inspection

2026-07-19

AI Summary

Although super-resolution reconstruction enhances image sharpness, it may inadvertently distort critical evidence required for defect detection, leading to uncontrolled false positive rates. This work proposes evaluating super-resolution methods not merely by reconstruction fidelity but by their capacity to preserve task-relevant evidence and enable operating point transfer between reconstruction and detection. A low-false-positive benchmark is established to decouple these two tasks. Through ten end-to-end experiments on both synthetic and real semiconductor mask data, bicubic interpolation, a local residual detector, a DeepLabV3-based direct detection model, and the jointly trained DPU-WaferSR are compared. Results reveal that learning-based reconstruction models, despite achieving high structural similarity, yield lower defect detection rates than simple interpolation; while direct detection consistently meets feasibility criteria, its recall for weak defects remains near zero, demonstrating that successful calibration does not guarantee effective real-world generalization.

Abstract

Super-resolution can make inspection images appear sharper without preserving the evidence needed to detect a defect. We study this failure mode with a benchmark that separates reconstruction from detection and evaluates both at a predeclared low false-positive rate. Ten end-to-end repetitions combine independently generated line/space and contact-hole images with model training, calibration, clean controls, weak defects, and a held-out defect morphology. Every reconstruction is scored by the same local residual detector, while direct and jointly trained detectors form a separate comparison track. Reconstruction fidelity and inspection utility diverge: the two learned reconstruction models attain the highest structural similarity yet detect fewer defect pixels than bicubic interpolation in every paired repetition. A direct DeepLabV3 detector reaches $0.1984\pm0.0385$ recall at $0.000174\pm0.000084$ false-positive rate and satisfies the held-out feasibility criterion in all ten repetitions. An illustrative joint model, DPU-WaferSR, passes independent clean calibration but exceeds the held-out limit in all ten repetitions, demonstrating that calibration success does not guarantee transfer. Weak-defect recall remains near zero for every feasible method. Applying the unchanged policies to 4,591 public Carinthia-S masks further reveals large method-dependent shifts on real SEM texture. These results support a simple conclusion: super-resolution for inspection should be judged by preserved task evidence and operating-point transfer, not reconstruction quality alone.

Problem

• super-resolution
• defect detection
• semiconductor inspection
• false-positive rate
• task evidence preservation

Innovation

• super-resolution
• defect detection
• low false-positive rate
• semiconductor inspection
• task evidence preservation