Research On Deep Learning-based Image Compression Technology

The boom in modern communication technology,as well as the rapid change in electronic and digital equipment,has placed greater demands on image compression efficiency,especially in the face of limited storage space and communication bandwidth.Traditional image compression methods have faced major bottlenecks in terms of further performance breakthroughs,and much of the emerging research has been devoted to exploring the use of deep learning to build high-performance end-to-end optimized image compression frameworks.However,existing end-to-end optimized image compression techniques still suffer from many shortcomings,such as pixel redundancy in the latent representations extracted by the encoder,excessive noise introduced by existing quantization techniques,and insufficient enhancement capabilities in the post-processing module.In response to these challenges,this paper has carried out research,which is summarized in three main areas:(1)To address the problem of pixel redundancy in the latent representations extracted by the encoder,this paper proposed the use of a hybrid domain attention mechanism embedded in the main encoder-decoder module and the entropy estimate module to reduce pixel redundancy in the latent representations by capturing the local and global correlations between image pixels.The hybrid domain attention mechanism includes a space-domain attention module and a channel-domain attention module,and during the end-to-end optimization process,the network generates weight vectors and space-domain mechanism masks through the channel-level attention and space-domain attention modules to guide the network in generating more compact features.(2)Aiming at the problem of excessive noise in quantization.This paper used a mixed quantization approach,i.e.,different quantization forms are used depending on the sensitivity of the different modules to the quantization form.Straight Through Estimation(STE)quantization for the main encoder-decoder part;and approximate quantization by adding uniform noise for the entropy estimate module.In addition,we proposed an inverse quantization module that learns the loss of floating-point numbers due to quantization and complements the quantization error.(3)To address the problem that the enhancement capability of the post-processing module is insufficient and can prone to fail when the network at the decoding end is sufficiently complex,we proposed a joint mean-aided information for post-processing enhancement.It is evident that the mean information is highly correlated with the latent representations,i.e.,the mean information contains a wealth of image information.On the basis of this theory,this paper proposed a post-processing enhancement network to complete the enhancement of image quality.In summary,this paper proposed three mechanisms to address the existing challenges of end-to-end optimized image compression,such as embedding hybrid attention,improving quantization,and post-processing using joint mean-assisted information,respectively.Experimental studies have shown that,on average,the three mechanisms deliver performance improvements of approximately 0.2 d B,0.18 d B,and 0.2 d B in PSNR,respectively,at BPP(average bits per pixel)equal to 0.2,and approximately 0.35 d B in combination.These experimental results demonstrate the validity of the relevant work in this paper.