Research On Key Techniques Of Image Compression Based On Significance Region

With the rapid development of computer networks and artificial intelligence,image processing technology,as an important method of multimedia data processing,has also made great progress.In order to adapt to changes in the network environment,image compression algorithms have also evolved image compression techniques for different scenarios.Image compression technology based on Region of Interest(ROI)is a technology that adopts a differential compression strategy for image content.With the advent of the 5G era,the rapid increase in the volume of multimedia data(images,video,audio,etc.)has put forward higher requirements for multimedia processing technology.Traditional commercial image compression methods have been difficult to adapt to the current application environment.When the wireless channel environment is relatively limited,people are more concerned about the sensory quality of the region of interest after reconstruction.The ROI-based image compression algorithm can better save the target of interest and improve the overall quality of the image.This paper proposes a compression algorithm based on deep learning(Deep Learning,DL)to solve this problem.The following is the research on this problem:The first thing to consider is the detection of saliency targets in the image.For different saliency target images,the background is separated from the saliency target,and then the saliency target region is set as the region of interest of the image.Prior to this,traditional ROI-based image compression algorithms used manual annotation of regions of interest to achieve image region separation.However,this method is too inefficient in large batch operations,and takes time and effort.With the rapid development of Convolutional Neural Networks(CNN),it has also brought new directions and discoveries to the field of salient object detection.The semantic segmentation and target detection algorithms that have appeared in the recent years are mostly based on Fully Convolutional Neural Networks(FCN).In this paper,the existing saliency detection algorithm can‘t combine high-level semantics and low-level features,in-depth research and improvement of CNN-based saliency detection network.In the image compression part,we propose an image compression framework based on Generative Adversarial Networks(GAN).In response to the problem of image compression and reconstruction at low bit rates,GAN can generate more delicate textures and detailed information for the image,making the reconstructed image have a higher subjective perception quality,and thus overcome the incomplete image of the traditional compression algorithm at low bit rate Severe distortion such as code block effect.In the network design process,since this article mainly discusses the image compression and reconstruction at a compression bit rate lower than 0.1bpp,this method changes the image compression bit rate by setting the number of channels in the bottleneck layer.In the design of the codec,we adopt a symmetric structure,and add 9 residual blocks to the decoder,which makes it easier to extract the high-level semantic information of the image during network training,thereby improving the sensory quality of the reconstructed image.In order to further improve the quality of image reconstruction at low bit rates,this paper also proposes a strategy of region-wise optimization.In the case of low bit rate,for the problem of poor image reconstruction and incomplete recovery,this paper uses different loss functions to perform network training on different regions of the image.In the background area of the reconstructed image,the network fills the reconstructed image with more textures and details to improve the subjective perceived quality of the image;while in the salient target area,the experiment will strictly preserve the true details and content distribution of the target,suppressing GAN Fill the content of the reconstructed image.Experiments prove that our method can achieve better compression of the region of interest compared with the existing ROI-based image compression algorithm JPEG2000,and exhibit better subjective perception quality.