Image Compression-oriented Fast Transform Algorithm And Its Application

In recent years,with the continuous upgrading of computers and smart devices,although the memory occupied by the acquired pictures is getting larger and larger,the capacity of the devices and the network transmission bandwidth are relatively slow to develop.How to reasonably store and effectively transmit images in the limited storage space and network bandwidth has always been a problem that researchers have to solve,and the generation of image compression technology can solve this problem.Image compression refers to the technique of removing excess information in an image and representing the original two-dimensional matrix information of the image with fewer bits.Although the traditional image compression algorithms compress the image to a certain extent,the compressed and reconstructed image has block effect and ringing effect,which cannot restore the image well.With the development of deep learning,a number of deep learning image compression algorithms have emerged.Although the compression effect is better than the traditional algorithm,the calculation time is longer when training the model,and the detailed information and texture information of the image still cannot be reproduced very well.In order to reduce the calculation complexity and calculation time in image compression,improve the image compression effect and reconstruct the image details,the main research contents of this thesis are as follows:(1)An image compression algorithm based on first-order moment two-dimensional discrete cosine transform is proposed.This algorithm designs a two-dimensional discrete cosine transform network structure of first-order moment without multiplication operation.First,the image is divided into 8×8 blocks,and then 4×4 blocks are extracted,and the first-order moment network structure is used to achieve lossy compression of the image.The algorithm only has an addition operation,and the calculation complexity is greatly reduced,which effectively improves the calculation efficiency of the algorithm.Compared with the Joint Photographic Experts Group(JPEG)and singular value compression(SVD)in MATLAB,The new algorithm has a simple structure,short delay,faster calculation time and speed than other algorithms,at the same times,the compression effect is better than other algorithms.(2)A deformable image compression framework based on multi-scale convolutional neural network is designed.In order to improve the compression quality of the image,the image is first subjected to the deformation processing operation.Through the image is processed by deformation,the details of the image are more obvious than the original image,and it is easier to further compress.The deformed image needs to pass a multi-scale convolutional neural network framework.The network framework consists of two convolutional neural networks and a traditional codec.The first convolutional neural network is a multi-scale compact representation convolutional neural network,which is used to learn the characteristics of the deformed image.After learning,the compact representation is sent to the traditional codec,and the decoded data is reconstructed by the second network which is the residual convolutional neural network restores the image.Through an optimization algorithm is used to optimize the parameters in the network,the network training time is reduced.By comparing with existing JPEG,JPEG2000,Web P,ARCNN algorithms,the image restoration effect is better than other algorithms.