Deep-learning-enabled geometric constraints and phase unwrapping for single-shot absolute 3D shape measurement (2001.01439v2)

Abstract: Fringe projection profilometry (FPP) is one of the most popular three-dimensional (3D) shape measurement techniques, and has becoming more prevalently adopted in intelligent manufacturing, defect detection and some other important applications. In FPP, how to efficiently recover the absolute phase has always been a great challenge. The stereo phase unwrapping (SPU) technologies based on geometric constraints can eliminate phase ambiguity without projecting any additional fringe patterns, which maximizes the efficiency of the retrieval of absolute phase. Inspired by the recent success of deep learning technologies for phase analysis, we demonstrate that deep learning can be an effective tool that organically unifies the phase retrieval, geometric constraints, and phase unwrapping steps into a comprehensive framework. Driven by extensive training dataset, the neutral network can gradually "learn" how to transfer one high-frequency fringe pattern into the "physically meaningful", and "most likely" absolute phase, instead of "step by step" as in convention approaches. Based on the properly trained framework, high-quality phase retrieval and robust phase ambiguity removal can be achieved based on only single-frame projection. Experimental results demonstrate that compared with traditional SPU, our method can more efficiently and stably unwrap the phase of dense fringe images in a larger measurement volume with fewer camera views. Limitations about the proposed approach are also discussed. We believe the proposed approach represents an important step forward in high-speed, high-accuracy, motion-artifacts-free absolute 3D shape measurement for complicated object from a single fringe pattern.

• The paper presents a deep learning framework that unifies phase retrieval, geometric constraints, and phase unwrapping for efficient absolute phase recovery.
• It leverages convolutional neural networks with dual-camera data to outperform traditional methods in both static and dynamic measurement scenarios.
• Quantitative evaluations on standard objects show high measurement accuracy, highlighting potential applications in high-speed industrial inspections.

Deep-learning-enabled Geometric Constraints and Phase Unwrapping for Single-shot Absolute 3D Shape Measurement

The paper presents an innovative approach that addresses existing challenges in Fringe Projection Profilometry (FPP) through deep learning techniques, specifically targeting the efficient recovery of absolute phase in single-shot 3D shape measurements. Recognizing the deficiencies in conventional methods, the authors propose a unified framework leveraging deep learning models to organically integrate phase retrieval, geometric constraints, and phase unwrapping.

The proposed system utilizes convolutional neural networks to process and interpret single-shot fringe patterns captured from dual-camera setups. The approach demonstrates how trained neural networks can learn, from vast training datasets, to predict "physically meaningful" absolute phases without the incremental steps typical in conventional methods. This advancement facilitates robust phase unwrapping of high-frequency fringe images across a larger measurement volume with reduced dependency on camera perspectives, enhancing the efficiency of shape measurement processes in high-speed scenarios.

Significant numerical results reveal that the authors' method competes favorably with traditional triple-camera stereo phase unwrapping (SPU), supplemented by adaptive depth constraint algorithms, in both static and dynamic measurement scenarios. The deep learning-based approach circumvents the inherent motion artifacts associated with multi-frame phase-shifting techniques, underscoring its potential for real-time applications.

The paper explores system calibration intricacies and quantitative assessments of measurement accuracy, further substantiating the reliability of their neural network approach. Error analyses on reconstructed standard spheres within the experimental setup highlight deviations that, while modest, affirm high-quality metric deliverables.

In conclusion, this research signifies an essential development in FPP by introducing deep learning's versatility to enhance the framework, optimizing computationally efficient and accurate 3D shape measurements. Future research avenues may explore integrating physical models into neural frameworks, facilitating comprehensive ecosystem solutions for optical profilometry, an endeavor propounded by the authors in their futuristic vision. However, among the limitations, the authors address potential setbacks in cases of inherent 2D image ambiguities, such as depth discontinuities, indicating room for further exploration and improvement using auxiliary data or models.

The implications are vast for industrial applications requiring high-speed, non-contact measurements—spanning sectors such as intelligent manufacturing and structural inspection—where reducing operational inefficiencies and computational overhead are critical. As such, these contributions meritoriously advance fringe projection methods, providing both theoretical enrichment and practical enhancements for AI-fueled optical imagers.