Research On Depth Video Coding Based On Convolutional Neural Network

3D video provides stereo perception by means of adding the number of viewpoints and depth data.As an important part of 3D video,depth video reflects the distance and distribution of objects in the scene.Depth video has significant differences compared with traditional color video,so it is of great importance to study more efficient depth video compression coding method.In recent years,the deep learning technology represented by convolutional neural networks(CNN)has developed rapidly and achieved fruitful results in the field of video processing and computer vision,which provide new ideas for further improving the performance of depth video coding.Based on the CNN technology and the characteristics of depth video,this thesis develops a depth video coding based on convolutional neural network.This thesis implements a convolutional neural network based variable resolution intra prediction method for depth video coding.Depth video consists of a large area of smooth or slowly changing regions and significant edges.By analyzing the content characteristics of depth video,this paper first downsamples the input depth video coding unit to reduce the resolution and compression space redundancy of the coding unit.Then,the low-resolution depth video coding unit is coded based on the conventional coding framework.Finally,the reconstructed coding unit is upsampled by the designed multi-level feature fusion network,which is able to recover the original size of the coding unit.Experimental results show that this method can effectively improve the performance of depth video coding.This thesis propose a color information guided variable resolution intra prediction method for depth video coding.In 3D-HEVC,color video and depth video are both included in each view.When the depth video is being encoded,the coding process of the corresponding color video has already accomplished.Based on the similarity between depth video and corresponding color video,a novel texture-assisted CNN is presented to handle the depth block up-sampling.The network is made up of several residual coding units(RCU)and the features of texture block are extracted to assist the reconstruction of the corresponding depth block.Experimental results show that the proposed method achieves competitive rate-distortion performance compared with the state-of-the-art approaches.