Deep neural network for processing single fringe patterns: Applications in phase recovery and denoising

This paper presents a Deep Neural Network Fringe Processor (DNNFP) for unified estimation of fringe orientation and frequency from single fringe patterns. Unlike conventional convolutional neural network (CNN) approaches that require prior image filtering and fixed input dimensions, our method processes raw fringe patterns of arbitrary size through a single neural network architecture. The DNNFP combines computational efficiency with operational simplicity, enabling consistent processing across diverse noise conditions without manual parameter tuning.

The proposed DNNFP analyzes local pixel neighborhoods to simultaneously extract orientation and frequency parameters, eliminating preprocessing steps and specialized algorithms. Experimental results demonstrate the method's practical utility in handling real-world fringe patterns, while qualitative analysis highlights advantages over CNN-based approaches in architectural simplicity and operational flexibility.

Key advantages include: (1) Elimination of preprocessing steps (denoising and reshaping) required by CNN methods, (2) single-network estimation of both orientation and frequency, and (3) parameter-free operation suitable for practical implementation. These features make the DNNFP particularly valuable for optical metrology applications where accurate phase recovery from single fringe patterns is essential. The method establishes a new framework for fringe pattern analysis that combines deep learning accuracy with practical usability.