Region Of Interest Coding Based On Deep Learning

With With the in-depth popularization and application of the Internet and big data,hundreds of millions of pictures and video information are generated and spread on the Internet every day,and the quality of pictures is getting higher and higher,gradually changing from high-definition to 2k,4k or even 8k.This has led to an increasing amount of image data,which puts tremendous pressure on data compression,transmission and storage,and also puts forward higher requirements for efficient image compression technology.The traditional general coding method has been difficult to meet the requirements,so the research on the coding of the region of interest is very meaningful.In the region of interest coding,the extraction of the region of interest(saliency region extraction)has always been a difficult point.The traditional region of interest extraction algorithm only regards the region with a certain characteristic as the region of interest.The extraction effect of the region of interest is very unsatisfactory.In response to this problem,this paper proposes a saliency region extraction algorithm based on fusion of multi-layer convolution features,which can extract regions of interest in complex pictures.In addition,in the region of interest extraction algorithm based on the fusion of multi-layer convolution features,we also proposed a region of interest coding algorithm,which can greatly improve the compression rate while ensuring subjective quality.Based on this,this article has done some work around the extraction and coding of regions of interest:(1)Based on the VGG16 network,a saliency region extraction algorithm based on the fusion of multi-layer convolution features is proposed.The specific method is to use the VGG16 network as the backbone network and add the BN layer to extract features.When extracting features,in order to supplement the shallow detail features,we downsample the shallow convolution features and fuse them into the deep convolution features.;Then the deep-level convolutional features are up-sampled and merged into the shallow convolutional features to guide the learning of features.Finally,when extracting the region of interest based on the features,in order to make full use of the features of each layer,we will extract the features The shallow features of is applied to the salient region extraction network.Experimental results show that this method can increase F-mea by 0.071 on average and MEA by 0.031 on average on the basis of the original network.(2)In terms of the training of the saliency region extraction network,we designed a layer-by-layer supervision method to guide the training of the network model,that is,the down-sampling map of the label map is used as the standard map of each deconvolution layer for supervised training.The experimental results show that the network trained by this method shows better robustness and generalization ability than other algorithms.(3)The idea of combining artificial intelligence methods and coding is proposed,and the saliency region extraction algorithm that combines multi-layer convolution features is used to obtain the region of interest in the image,and divide the image into the region of interest and the background region,We adopt different quantification standards for different regions of the image,the quantification of the region of interest is more detailed,and the quantification of the non-interest region is coarse quantification.The experimental results show that this coding method can not only increase the average compression rate to 82..89%,but also ensure that the distortion rate of the region of interest reaches the lowest.