Neural Network Model for Binocular Disparity Extraction

摘要

In the broadcast media, the utilization of depth information from two cameras is important. In order to display 3D information, the reconstructed 3D images should look as natural as the real world even if the images are from two points only. For this purpose, it is necessary to reconstruct 3D information as in the human brain. The paper proposes a neural network model which can extract binocular disparity accurately by the use of edge information. It is shown by simulation that the proposed model can specify relative positions of edges in 3D space without suffering from the false target problem. A 4th layer is specially added to improve the selectivity of disparity and the formation of iso-parallax contors. It is shown that the 4th layer contributes much to the improvement of the resolution in depth.