Image Compression Based On Generative Adversarial Netwroks

With the growth of information technology,huge image data and severe resource constraints have brought great challenges to the development of multimedia.Image compression technology can reduce the redundant information in image,thus reducing the pressure of storage and transmission.Recently,deep learning has attracted wide attention in the field of computer vision.Neural network is utilized to extract the deep features of images,and its application in image classification,target detection and image super-resolution has achieved great breakthrough.Although,it is still in the preliminary stage of research in image compression,some existing studies show that it has huge research potential and value.This paper discusses the loss function,the comprehensive evaluation of restored image quality and the data dependence based on the generative adversarial network.The main research contents and conclusions of this paper are as follows:(1)Aiming at the problem of data dependence in the generative compression algorithm,an improved method based on spectral normalization is proposed.By analyzing the optimization space of the discriminator,the design idea of the improved discriminator is expounded and the constraints required for improvement are analyzed.The experimental results show that the proposed algorithm can improve the diversity and clarity of the restoration map significantly compared with the general generative countermeasure network.(2)To solve the problem that the random generated images are different from the original images,an improved method based on mutual information maximization is proposed.The mutual information between compressed data and restored images is taken as an additional optimization objective to complete the compression of multi-class images.The experimental results show the image restoration ability of this algorithm on single-class image and multi-class image data sets.Compared with other image compression algorithms,the superiority of this algorithm is verified.(3)This paper introduces two kinds of loss functions commonly used in neural network based image processing technology,including pixel-level loss function and feature extraction loss function.A hybrid loss function is proposed and applied to the network training of this algorithm.The optimal weight allocation scheme is found through the experimental results,and the loss function proposed in this paper is compared with other losses.The experimental results of the function are compared and analyzed.(4)By analyzing the common evaluation methods of subjective and objective image quality,a set of reasonable restoration map quality evaluation scheme is formulated according to the test dataset.And different comprehensive evaluation methods are proposed for different types of data sets in the experiment,so that the comprehensive index may better represent the restoration performance of our algorithm.