Analyzing and Improving of Neural Networks used in Stereo Calibration

Abstract

In this paper, CCD cameras are calibrated implicitly using BP neural network by means of its ability to fit the complicated nonlinear mapping relation. Dense sample data is acquired by using high precisely numerical control platform, and the variances error (PVE) is adopted during training the neural network. The error percentages obtained from our set-up are limitedly better than those obtained through mean square error (MSE). The system is generalization enough for most machine-vision applications and the calibrated system can reach acceptable precision of 3D measurement standard. It is expected that, with this approach, we can maintain the major advantage of linear methods and obtain improved accuracy without any complicated mathematical modeling process thank to nonlinear learning capability of neural networks. The value p needs to be decided by experiments, and the reconstruction images will be distorted if the value is more than 6.