The elimination of errors caused by shadow in fringe projection profilometry based on deep learning

The fringe projection profilometry (FPP) has been regarded as a classical and mature technique for 3D shape measurement. However, in practical applications, shadow is un-avoided in the imaging process and causes errors in many FPP systems. In this paper, the errors in FPP systems caused by shadow are first analyzed. Then, a direction-aware spatial context module based network is proposed for detecting the shadow regions of a fringe image. Further, a repairment method is developed based on a generative adversarial network combining some simple processes. The training datasets are rendered by a graphic software to easy the training of the networks. The proposed method can repair the shadow regions successfully with only one fringe image and so it can be applied in varieties of FPP systems. The feasibility and the accuracy of improved by the proposed method have been illustrated by abundant experiments.