AI-Generated Per-Measurement References for Explainable Destriping of Industrial Line-Scan Metrology Images

Horizontal banding caused by illumination or sensor-response non-uniformity is a recurring problem in stitched line-scan measurements. Direct generative restoration can remove the visible bands, but it cannot by itself provide a traceable measurement because it may redraw specimen details. This study evaluates a hybrid design in which a local image-editing model creates a separate approximate clean reference for every raw measurement. The generated image is never returned as measurement data. Instead, its scanline intensity distributions are paired with those of the raw image, a robust quadratic response model is fitted for each scanline, and the model is inverted on the original high-bit-depth samples. A second deterministic pass estimates spatially varying periodic residual gain on a 9x9 grid and analytically composes it with the first model. On two available wafer scans, the periodic log-domain proxy was reduced by 99.31% and 98.30%, respectively. These results establish technical feasibility, but the small data set and absence of calibrated clean ground truth do not yet establish metrological accuracy.