Projector distortion correction in 3D shape measurement using structured-light system by deep neural networks

In structured light system, lens distortion of the camera and projector is the main source of 3D measurement error. In this letter a new approach of using deep neural networks to address this problem is proposed. The neural network is with one input layer, five densely-connected hidden layers and one output layer. A ceramic plate with flatness less than 0.005 mm is used to acquire the training, validation and test data sets for the network. It is shown that the measurement accuracy can be enhanced to 0.0165 mm in RMS value by this technique, which is an improvement of 93.52%. It is also verified that the constructed neural network is with strong robustness.