A unified deep learning framework for phase demodulation in off-axis digital holographic interferometry: applications in polymer fiber characterization and 3D imaging

Abstract Off-axis digital holographic interferometry is a powerful non-destructive technique for quantitative phase imaging, yet conventional reconstruction methods face substantial computational challenges when processing complex experimental scenarios. This study presents a developed convolutional neural network framework for direct phase demodulation from off-axis holographic interference patterns, eliminating the need for traditional multi-step reconstruction procedures. This suggested method advances real-time opto-mechanical experiments without requiring explicit knowledge of propagation distances between focus and defocus conditions. The network architecture incorporates multi-scale encoder pathways, focus-adaptive processing, and distance-agnostic decoding to handle four primary holographic scenarios: ordinary off-axis holograms, axially tilted samples with mixed focus conditions, angularly misaligned specimens, and combined multi-distance and angular variations. The framework was trained and validated using experimental 1000 high-resolution holographic patterns from diverse polymeric fiber specimens. Experimental validation demonstrated exceptional performance with validation PSNR of 25.66 dB and SSIM of 0.903, while reducing processing time from 9–16 min to 30–48 s per image. The network eliminates the need for explicit knowledge of propagation distances, focus conditions, or object orientations. Practical applications include dynamic opto-mechanical characterization of isotactic polypropylene fibers under mechanical drawing and three-dimensional morphological analysis of iPP/TiO₂ nanocomposite fibers through holographic tomography. This work advances holographic metrology by providing a unified solution for quantitative phase imaging under the four experimental configurations.