Error compensation for phase retrieval in deflectometry based on deep learning

Abstract

Fringe patterns are widely applied in optical metrology, and phase retrieval is an essential process for decoding surface information. In the field of phase measurement deflectometry (PMD), phase errors in the phase retrieval process have more significant effects for PMD is a slope-based technique and is more sensitive to low-frequency errors. The main factors affecting the quality of the captured fringe images include the gamma effect of the liquid crystal display (LCD) screen, the random noise from the charge-coupled device (CCD) camera, and the random noise amplified by the defocused fringe patterns. Conventional methods compensated the phase errors of these factors separately with different methods, which are inefficient in handling the errors from coupling factors effectively. In this paper, we propose a deep neural network to compensate for the phase errors resulting from the combination of the factors. Experimental results demonstrate that the proposed network can significantly suppress the errors in phase retrieval with non-ideal fringe images. The phase errors can be reduced in both simulated and authentic data for deflectometry, which verifies the robustness and effectiveness of the proposed method.