Research On Image Compression Algorithm Based On Nonlinear Transformation

In recent years,digital image compression has always been a hot issue in digital communication technology.Its core purpose is to reduce the amount of data needed to represent digital image on the basis of ensuring image quality,but it is often impossible to have both.The effect of traditional image compression methods in low bit rate is not satisfactory,and there are some problems such as block effect and texture blurring.At present,deep learning has achieved great success in image recognition,vehicle detection and other fields,some of which even surpass the traditional methods,but there will still be texture blur in ultra-low bit rate.Therefore,this thesis will make use of the advantages of deep learning and traditional image processing methods in image compression to further study the quality improvement of image compression and reconstruction at low bit rate or even ultra-low bit rate.1.This thesis designs an image compression arithmetic based on Variational Auto-Encoder(VAE),which is suitable for image compression coding at low bit rate and ultra-low bit rate.First,according to the multi-resolution characteristics of wavelet transform,a deep learning compression arithmetic model that can efficiently compress low-frequency images after wavelet transform is trained.Then,the reversibility of wavelet transform is used to reconstruct the image to solve the problems of false contour and block effect in the image at low bit rate.In particular,a super prior network is added to assist in image reconstruction.Experimental results show that this method is superior to traditional methods and methods based on deep learning in terms of image reconstruction quality.2.On the basis of the above ideas,an image compression arithmetic based on Convolutional Auto-Encoder(CAE)is proposed,which is suitable for image compression coding at low bit rates.The difference from the traditional image compression arithmetic is that the input object of the traditional image compression arithmetic is a natural image,and the input object of the arithmetic is a residual block filtered out of the natural image through a series of processing.In particular,the data distribution of the residual block and the potential representation of redundant information are studied,and the radial Gaussianization(RG)is used to map it to a spherical Gaussian distribution,thereby reducing the local statistical dependence of the data distribution.Meanwhile,using the new optimizer Ranger in the compression algorithm can better accelerate its training convergence.Experiments show that compared with traditional image compression methods and image compression methods based on deep learning,the method proposed in this thesis has a good performance on PSNR and SSIM.