Fringe pattern denoising with a deep neural network

Interferometric images (e.g., fringe and wrapped-phase patterns) are pivotal in optical metrology and industrial inspection, yet their accuracy is often compromised by noise from sensor imperfections, environmental fluctuations, and instrumental limitations. While many traditional denoising methods require case-specific parameter tuning or prior fringe orientation estimation, deep learning approaches face challenges in accessibility and implementation complexity. We propose a Deep Neural Network Filter (DNNF), a novel denoising technique that combines the robustness of deep learning with the simplicity of convolution filters. Our method employs a custom-designed deep neural network (DNN) trained on a synthetic dataset encompassing diverse local fringe structures (orientation, frequency, curvature) and noise profiles (Gaussian phase noise). The DNNF operates as a parameter-free, plug-and-play solution, processing images analogously to a convolution filter without requiring manual adjustments or fringe orientation estimation. Extensive experiments on synthetic and real interferometric images demonstrate the DNNF's competitive performance against some state-of-the-art methods. Key achievements include: good noise suppression while avoiding artifacts common in anisotropic filters. Balanced processing time without compromising accuracy. Consistent performance across varied fringe densities, orientations, and noise levels, validated in moiré interferometry, Michelson interferometry, and Electronic Speckle Pattern Interferometry (ESPI) applications. The DNNF bridges the gap between advanced deep learning and practical usability, eliminating the need for expertise-dependent parameter tuning. Its efficiency, robustness, and simplicity make it a promising tool for real-world interferometric image processing, particularly in scenarios where traditional methods falter.