Phase extraction and denoising through deep learning-based interference pattern analysis

Digital Holographic Speckle Pattern Interferometry (DHSPI) is a custom-made portable coherent interferometry system for on-field measurements, enabling precise surface displacement detection with high-quality interferograms under practical conditions. As a portable coherent interferometry technique, it suffers from measurement noise, which can significantly impacts downstream machine learning processing and automated analysis workflows, while systematic noise from small motions and environmental vibrations can compromise measurement consistency across extended acquisition sequences. This work presents a neural network-based approach that simultaneously addresses noise suppression and acquisition efficiency to minimize systematic drift effects. To this end, EffiSegNet is used, an asymmetric convolutional neural network architecture, which was trained on 10,000 synthetic images incorporating realistic noise models including photon shot noise, electronic read noise, and quantization effects. The system demonstrates dual capabilities: effective denoising of wrapped phase maps achieving 21.04 dB PSNR with 88.77% SSIM, and wrapped phase extraction from reduced interference sequences. The network maintains 95% of conventional five-step phase shifting performance when processing just three interference images, reducing data acquisition by 70% while preserving measurement fidelity. These findings establish neural networks as viable solutions for next-generation DHSPI systems requiring both enhanced noise robustness and accelerated measurement capabilities for industrial deployment.