YASO

Abstract

This paper presents a lossy image compression methodbased on neural network. Our architecture just consists ofa recurrent neural network (RNN)-based encoder and decoder, a binarizer. This paper makes contributions in thefollowing two aspects: 1) Preprocess the input images sothat the encoder could work on images with arbitrary size;2) Optimize the number of output channels of the binarizer, and our method ensures that the compressed image usesless than 0.15 bpp; 3) Our network is suitable for highresolutionimages thanks to a sub-pixel architecture. As aconsequence, we find that the optimized method generallyexhibits better rate-distortion performance than standardJPEG compression methods on the Kodak dataset.