Variable Rate Image Compression with Recurrent Neural Networks

A large percentage of Internet traffic is driven by requests from mobile devices with relatively small screens and often stringent bandwidth requirements. For websites that strive to be as responsive as possible, it is crucial to transmit image previews as fast as possible in the form of thumbnails. As a result, any improvement to thumbnail compression will significantly enhance the experience of mobile device users. We propose a general framework for variable-rate image compression and a novel architecture based on convolutional and deconvolutional LSTM recurrent networks. Our models address the main issues that have prevented autoencoder neural networks from competing with existing image compression algorithms: (1) our neural network only needs to be trained once regardless of the input image (it is not trained per image), its dimensions, and the desired compression rate; (2) the algorithms we describe are progressive, meaning that the more bits are sent, the more accurate the reconstruction of the image is; and (3) the proposed architecture is at least as efficient as a purpose-trained autoencoder for a given number of bits. On a large-scale benchmark of 32$\times$32 thumbnails, our LSTM-based approaches provide better visual quality than (headerless) JPEG, JPEG2000 and WebP, with a storage size that is reduced by 10% or more.