Depth-image Coding Using Neural Networks for 3D Video Transmission
Yih-Chuan Lin, Pu Jian Hsu · 2013
In this paper, neural networks are used to classify the content of depth images of 3D videos into three kinds of content classes with distinct characteristics in order to reduce coded data-rate. One type of neural network is flrst employed to identify the boundary information across the foreground and background objects based on the information from the video texture frame and the associated depth image. Another type of hybrid neural network is trained to flt the surface of depth values over the foreground objects. The similarity between successive depth-image frames is captured by using one other type of neural network to reduce the amount of data required to deliver them. With experimental results, the transmission bit-rate and frame reconstruction quality are evaluated by comparisons to that of using H.264/AVC codec on the same depth image sequence.