Neural-Network Model for Compensation of Lens Distortion in Camera Calibration

Camera calibration for machine vision is critical in three-dimensional (3-D) measurement systems based on a digital light processing (DLP) projector and a camera. The Z-height of the measurement point is calculated using the phase value observed by the camera when a fringe pattern is scanned from a projection onto an object. On the other hand, the X and Y coordinates are obtained from the camera coordinates using a transformation matrix, and the mathematical model for lens distortion is additionally used. However, the errors for x and y coordinates are 10 times larger than the z-height error in an experiment. This is because the lens distortion is not sufficiently compensated in the mathematical model considering only the position from the lens center. Therefore, the neural network (NN) model that considers the measurement distance in addition to the position is proposed in this paper. Experiments were conducted on a 100 x 100 mm(2) area, and a maximum error of 0.5 mm is observed for the mathematical model. However, when the NN model considering the height of the object is used, the error is reduced by 60% to 0.2 mm.