Abstract

Fringe Projection Profilometry (FPP) is widely used in shape retrieval for non-contact and high-speed three-dimensional (3D) measurements. Phase unwrapping is critical to FPP. Among all the proposed FPP systems, the single-camera and single-projector system projecting only single-frequency patterns is the most ideal one in terms of 3D imaging speed and system cost. However, this type of system needs robust spatial phase unwrapping (SPU) methods. A well-known problem with SPU methods is that the unwrapped result is path-dependent. Although quality-guided phase unwrapping methods achieve good results in reducing the path dependence by excluding invalid points from the unwrapping process, they still face great challenges in dealing with phase maps with phase discontinuity and motion blur. This study proposed a learningbased invalid points detection approach for FPP. A deep convolutional neural network (DCNN) is suggested to learn the discriminative features of points from the composite map of background intensity, intensity modulation, and phase map. After training, this DCNN is able to classify each point in phase maps into three classes. The training data is collected using a hand-held FPP system, which achieves 3D reconstruction via temporal phase unwrapping. The labeling of training data is automatically accomplished through a registration-orientated labeling procedure. Experiments on a large-scale dataset of real objects demonstrate that complex phase maps with motion blur and phase discontinuity can be unwrapped correctly by the flood-fill algorithm after the invalid points are detected by the proposed approach.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call