PEENet for phase unwrapping in fringe projection profilometry

Abstract

In this paper, we propose a lightweight deep convolution neural network, named PEENet, for high resolution image phase unwrapping in fringe projection profilometry on the device with limited performance. In our method, the dilated convolution strategy is applied to the networks, which increases the receptive field of the network while reducing the amount of network parameters. In the PEENet, we use Atrous Spatial Pyramid Pooling (ASPP) structure which can reduce the network parameters (total 0.48 million) while can extract the deep features of image. We also use Edge-Enhanced Block (EEB) structure, which can enhance the edge features of the image. We conducted ablation experiments to explore the effect of different network structures on network performance and then we compare our method with the other lightweight deep convolution neural network with the same training and testing datasets. We also build a new dataset that contain more different situations which can enhance the generalization ability of the network. The results show that our method achieves higher accuracy with fewer parameters and the new dataset works well.