Robust Shape-from-Focus via Physics-Inspired Distortion-Aware Focal Depth Regression

Shape-from-Focus (SFF) is attractive for microscopic three-dimensional measurement, but high dynamic range (HDR) surfaces and weak-textured surfaces distort the focus curve through saturation, spurious peaks, and low signal-to-noise ratios. These distortions violate the unimodal assumption used by Gaussian peak localization and limit post-processing-only correction. This paper proposes a physics-guided distortion-aware SFF pipeline for opaque single-surface targets. The Distortion-Aware Focal Depth Regression Network (DAFDR-Net) learns from synthetic focus-curve distortions and uses Channel-wise Feature Attention (CFA) and Soft Peak Localization to reweight distortion-sensitive temporal-response features while preserving a peak-localization prior. Its foreground validity output is further used for confidence-guided adaptive smoothing. On an HDR free-form surface dataset, the proposed pipeline reduces RMSE by 36.5% relative to an MRF optimization method and compresses the 99th-percentile absolute error from 0.181 to 0.033. On weak-textured monocrystalline silicon wafer data, it reduces flat-region depth standard deviation by 51.3%.