UHRNet: a deep learning-based method for accurate 3D reconstruction from a single fringe-pattern

Abstract

ABSTRACT The quick and accurate retrieval of an object’s height from a single fringe-pattern in Fringe Projection Profilometry has been a topic of ongoing research. While existing single-shot fringe-to-depth CNN methods can directly generate height map from one pattern, their accuracy lags behind traditional phase-shifting techniques. To improve accuracy, we propose a U-shaped High-resolution Network (UHRNet). The network utilizes U-Net’s encoding-decoding structure as the backbone, employing Multi-Level Conv Blocks and High-resolution Fusion Blocks to extract features. Additionally, a compound loss function, combining Structural Similarity Index Measure Loss (SSIMLoss) and chunked L2 loss function, is devised to enhance the details of 3D reconstruction. Our method has been experimentally proven to be efficient with an average RMSE of 0.443 mm, which is 64.67% lower than hNet and 33.31% lower than ResUNet. These results indicate that the proposed approach significantly enhances the accuracy of 3D reconstruction from a single fringe-pattern.