Deep-Learning Based Three Channel Defocused Projection Profilometry

Speeding up the process of fringe projection profilometry to make it more suitable for dynamic three-dimensional measurement is one of the targets of structured light research. In this work, based on the fact that the frame rate of the projector can be much higher than that of the camera, we propose a three-channel binary defocused projection method to break through the speed bottleneck of the camera. Three sequential phase-shifting binary patterns are projected in a defocused way by the three color channels of the projector respectively during one camera’s exposure time. To tackle the issue of color crosstalk, a deep-learning based end-to-end fringe rectification method is first introduced. As a result, only three camera shots are required for single scene reconstruction. Through experiments in three static scenes and one dynamic scene, we demonstrate that our method substantially speeds up the reconstruction rate with an acceptable range of accuracy loss. In addition, given the limited available datasets in the fringe projection profilometry field, a new dataset is also released.