Research And Application Of Image Compression Algorithm Based On Generative Adversarial Network

In recent years,image sharing has become one of the main channels for socializing on the Internet.For example,Weibo has 184 million daily users,and more than 100 million photos are uploaded and exchanged via Weibo every day.The storage space required for pictures also increases rapidly with the increasing pixel size of camera equipment such as mobile phones.Therefore,image compression plays a vital role in ensuring low-cost storage and sharing of the entire Internet.To reduce the storage space as much as possible,the encoding bit rate needs to be greatly reduced.However,the traditional engineering compression algorithm has unsatisfactory visual effects at low bit rates,blurring,blockiness and even lack of color.At present,deep learning has achieved good results in the fields of object detection,tracking,classification,etc.Some have even surpassed the traditional methods.However,in the direction of image compression,deep learning related technologies have not been fully researched.The adversarial generative network proposed by Goodfellow and others in 2014 brought us new ideas in image generation and superresolution due to its good scalability and clever weak supervised learning model.These image generation technologies also bring new thinking to the field of image compression.This article will use the advantages of deep learning and adversarial generative networks in image processing and image generation to conduct in-depth research on the improvement of image compression quality at low bit rates(less than 1bpp)or even ultra-low bit rates(less than 0.1bpp).In the field of deep learning,the extraction of image features cannot be separated from the convolutional neural network.First,this paper designs an autoencoder that includes an encoding end and a decoding end,which is based on a convolutional neural network.At the encoding end,the original image gradually reduces the spatial scale of the image through the convolutional layer to obtain the image features,and then reduces the information redundancy in the image by a quantizer to form a bit stream for transmission to achieve the purpose of image compression;at the decoding end,Re-encode the bitstream,gradually restore the scale and texture through the convolutional layer,and finally reconstruct the image.On this basis,this paper also designs a multi-scale ‘transcendental'mechanism to improve the rate distortion performance and better learn the texture details in the image.Experiments show the effect of the multi-scale‘transcendental'mechanism.And at low bitrates,our method goes beyond traditional image compression methods.Then,this paper proposes the introduction of adversarial loss based on adversarial generative network for‘rate-distortion'optimization in the case of blurred images,ultrasmooth textures,poor subjective visual effects and complete failure of traditional evaluation indicators at ultra-low bit rates.On the basis of the multi-scale autoencoder we proposed,we added the multi-scale discriminator used in the field of super-resolution image generation,and proposed a loss function closer to the perceived similarity.The distortion objective function introduces an adversarial loss function for end-to-end training.Experiments show that the image reconstructed by our method is significantly better than the reconstructed image obtained by the best traditional image compression method at the current low bit rate.