Universal phase retrieval transformer for single-pattern structured light three-dimensional imaging

Abstract

Deep learning-based single-pattern phase retrieval methods for structured light three-dimensional imaging have shown significant performance improvement over traditional methods. However, these methods are typically designed for datasets containing patterns with a particular frequency and coding direction. To address this limitation, we propose a universal phase retrieval transformer (UPRT) for single-pattern structured light three-dimensional imaging, which can handle patterns across various frequencies and coding directions. Specifically, we propose a two-dimensional frequency filtering block by analyzing a single pattern from the perspective of the frequency domain, allowing adaptive extraction of frequency components. Additionally, a line-based frequency processing block is proposed to capture both vertical and horizontal features in the frequency domain, and a simple yet effective frequency processing method is designed to achieve information interactions within lines and channels for this block. Experiments on single-pattern phase retrieval demonstrate that UPRT achieves state-of-the-art performance, with a 10.3% improvement over traditional methods. Furthermore, the proposed UPRT achieves robust results across patterns with diverse frequencies and coding directions, demonstrating its strong generalization abilities. Source code is avaliable at https://github.com/jianwensong/UPRT.