Synthetic training datasets generating for fringe projection profilometry based on deep learning

Abstract

Although Deep learning-based Fringe projection profilometry (FPP ) has some success in in three-dimensional(3D) shape measurement,it still exists some difficult problems,especially, collecting a huge of training datasets is particularly troublesome and inconvenient.We introduce Blender software to simulate virtual scanning in 3D shape measurement, present a synthetic training datasets generating method.The proposed method can produce an effective training datasets that does not need real-world 3D scanning procedure.It can automatically produce deformed fringe patterns of the tested object in a very short time when a group of phase-shifting fringe pattern is projected onto the tested object in virtual scanning way,then calculate the wrapped phase using phase-shifting demodulation algorithm and obtain the 3D shape information from the continuous phase after the measurement system is calibrated.To verify the effectiveness of our method, we did simulations and real experiments. The results show that our method is effective for the measuring the complex object’s shape. Compared the training datasets obtained with real-world scanning,the network model trained by training datasets obtained from our proposed method has the similar accuracy and generalization ability, but our method is simple and fast in preparing training datasets for network.