Research On 3D-AVS Prediction Techniques And Feature MAP Compression Techniques

In recent years,3D video has become an indispensable part of people's daily lives.AVS working group(The digital Audio and Video codec technical Standards working group)completed the 3D-AVS2 standard in 2014.The 3D-AVS2 video coding standard uses a texture map and a depth map coding method.The encoding of texture maps and depth maps of each viewpoint adopts the AVS2 encoding method,coupled with the intervision prediction technique and the improvement of the encoding of the depth map,forming a 3D-AVS2 encoding method.This paper has finished two parts of the current video standard 3D-AVS2.The first part proposed two optimization solutions for the memory usage problem of the motion information storage unit for the requirements of the domestic 3D-AVS Working Group,t.The first optimization scheme changed the prediction unit to motion information storage unit and changed backward(second)reference and backward(second)motion vector to 8bits to storage.Compared with RFD11.1,there are 0.66% and 2.69% gains under LDP and RA respectively in synthetic viewpoint.The second optimization scheme changed the prediction unit to the motion information storage unit and saved the backward(second)reference.The backward(second)reference is calculated by forward(first)motion vector,the distance between the forward(first)reference number and the current frame number,the distance between the backward(second)reference number and the current frame number.Compared with RFD11.1,there are 0.66% and 0.88% gains under LDP and RA respectively.For dependent view's depth maps of 3D-AVS2,the second part proposed a fast encoding method--depth map fast CU partition algorithm for depth maps of 3D-AVS2.This algorithm calculates the histogram of the CU first and then calculate the length of interval of gray value distribution.If the length is less than the threshold Th L,stop splitting the current CU.Otherwise,continue dividing the current CU.Compared with RFD11.0,this method reduces 0.04% encoding performance and saves 10.77% encoding time under LDP in synthetic viewpoint.And there is 0.10% loss and saves 15.99% encoding time under RA.In 2017,Facebook released Faiss,a library of algorithms for quickly searching similar multimedia files based on features.To use Faiss for image retrieval,you first need to extract features.The features extracted by the convolutional neural network can effectively mine image information.However,the high-definition image feature map is very large,which brings a large amount of computation for image feature retrieval.In this paper,the VGG network is used to extract the feature map before fully connected layer.The feature map is reconstructed using the VLAD algorithm,and then the features are quantized and coded.The decompressed feature vector is input into Faiss.It is concluded that the quantitative multiplication rate is inversely correlated with the accuracy,and the range of the decrease in the search accuracy rate can be tolerated in the case where the quantization loss is not large.On the whole,the proposed 3D-AVS2 inter-view motion parameter inheritance technique modification scheme has a certain performance gain compared to the current coding standard;the depth map fast CU algorithm saves the coding time when the performance loss is not large.The study on compression of feature maps shows that the retrieval accuracy is high when the quantization loss is small.