FDCGAN: frequency domain constrained generative adversarial network for overexposed fringe image restoration

Restoring overexposed fringe images in fringe projection profilometry (FPP) is a challenging task due to severe saturation and loss of fine structural details, which often renders traditional methods ineffective. In this work, we propose FDCGAN, a Frequency Domain Constrained Generative Adversarial Network designed to address this issue by combining spatial and frequency-domain learning. The proposed framework embodies a structurally coherent design that couples hierarchical feature encoding with multi-scale adversarial supervision, enabling both global structure preservation and local detail recovery, while normalization strategies regulate information flow and enhance resilience to photometric degradation. To further enhance restoration quality, we introduce a set of frequency-aware learning losses, including Fourier magnitude loss and high-frequency loss, which guide the model in reconstructing realistic fringe patterns. Experimental results demonstrate that FDCGAN significantly outperforms existing methods, especially under extreme overexposure, where fringe patterns and object shapes are barely visible. Even in such cases, FDCGAN can infer plausible fringe structures that significantly enhance the quality of 3D reconstruction. These findings highlight FDCGAN's robustness and its potential for improving measurement reliability in real-world FPP systems under challenging lighting conditions.