HRP-Net: an end-to-end deep neural network for high-reflection profilometry

Fringe projection profilometry for surfaces with high reflection material is always a challenging task because the captured fringes often suffer from over-exposure with truncated intensities. To address this problem, HRP-Net, an end-to-end deep learning framework, is proposed in this paper with two modules. On the one hand, the dense residual feature extraction module (DRFEM) is proposed to extract the fringe, where the over-exposure is suppressed while normal components are reserved. On the other hand, based on the extracted features, the depth estimation module (DEM) is proposed to predict the depth map. In DEM, a fused attention module (FAM) is used to fuse spatial and channel-wise features, and a multi-scale atrous convolution block (MSAC) are introduced to effectively utilize multi-scale features. Experimental results demonstrate that HRP-Net significantly improves depth estimation accuracy and robustness in reflective scenarios, effectively reducing depth discontinuities and preserving structural details compared to existing network-based methods.