Similar Image Coding Algorithm Based On Regression Models

With the development of the Internet and the popularity of cameras and mobile phones,a large number of pictures are uploaded to major social platforms every day,and the number of pictures is growing geometrically.In addition,with the update of devices,the size of photo is getting bigger,naturally it demands huger storage of devices.Traditional image coding methods,such as JPEG and JPEG 2000,compress a single image by removing the redundant information inside the image.However,these methods ignore the redundant information between images.In order to make use of the redundant information between images and improve the efficiency of image coding further,this thesis proposed a similar image coding algorithm based on regression models.Different from the existing image set coding algorithm using local features,the purpose of this thesis is to independently encode a single image.For each image to be encoded,some reference pictures that are most similar to it are searched from a huge database,then we mimic the three-step prediction in image set coding algorithm using local features,geometric transformation and photometric transformation are used to process reference pictures,and finally the image sequence is encoded using HEVC inter prediction.The thesis has done the following works: firstly,we devised a method to build a reference pictures database and use image retrieval based on deep learning to determine reference pictures.Then a geometric transformation method based on ridge regression has been proposed,rather than using traditional SIFT features,it extracts and matches the Affine-SIFT features between reference picture and image to be encoded.After that,we can get the optimal parameters via ridge regression and finally get the affine transformation matrix.Moreover,the thesis introduces a geometric transformation method based on Convolutional Neural Networks(CNN)and apply hierarchical geometric transformation to images,first perform affine transformation on a reference picture,and continue to perform TPS transformation on the deformed image.Finally,we make an improvement in traditional photometric transformation of reference pictures.The luminosity correlation between reference picture and image to be encoded is not regard as a linear model,by using Support Vector Regression(SVR)and introducing a kernel function,we implement nonlinear photometric transformation.It makes a less luminosity difference between reference picture and image to be encoded.Experimental simulation was performed on six image data sets,and the experimental results show that our similar image coding algorithm based on regression models improves the efficiency of image coding distinctly.In all six image data sets,at the same bit rate,similar image coding using geometric transformation based on ridge regression improves the PSNR approximately 0.1~2.1dB compared with HEVC intra coding,similar image coding using geometric transformation based on CNN improves the PSNR approximately 0.3~1.7dB compared with HEVC intra coding,similar image coding using photometric transformation based on SVR improve the PSNR approximately 0.1~0.3dB compared with similar image coding using traditional linear photometric transformation.