Full Resolution Image Compression with Recurrent Neural Networks

Abstract

This paper presents a set of full-resolution lossy image compression methods based on neural networks. Each of the architectures we describe can provide variable compression rates during deployment without requiring retraining of the network: each network need only be trained once. All of our architectures consist of a recurrent neural network (RNN)-based encoder and decoder, a binarizer, and a neural network for entropy coding. We compare RNN types (LSTM, associative LSTM) and introduce a new hybrid of GRU and ResNet. We also study "one-shot" versus additive reconstruction architectures and introduce a new scaled-additive framework. We compare to previous work, showing improvements of 4.3%-8.8% AUC (area under the rate-distortion curve), depending on the perceptual metric used. As far as we know, this is the first neural network architecture that is able to outperform JPEG at image compression across most bitrates on the rate-distortion curve on the Kodak dataset images, with and without the aid of entropy coding.

Deep Multiple Description Coding by Learning Scalar Quantization

“…Like [29], soft relaxation of quantization is used to resolve the nondifferentiability problem of the quantization function [27]. Different from [27,29], the thumbnail images are compressed by a recurrent neural networks architecture, in which a stochastic rounding operation makes feature maps binarized [30]. Recently, a virtual codec network has been learned to imitate the projection from the represented vectors to the decoded images to make the image compression framework trainable in an end-to-end way [11].…”

Variable Rate Deep Image Compression With Modulated Autoencoder

“…In deep image compression [1], [7], [8], the handcrafted analysis and synthesis transforms are replaced by the encoder z = f (x; θ) and decoderx = g (ẑ; φ) of a convolutional autoencoder, parametrized by θ and φ. The fundamental difference is that the transforms are not designed but learned from training data.…”

“…The problem is solved using gradient descent and backpropagation [18]. To make the model differentiable, which is required to apply backpropagation, during training the quantizer is replaced by a differentiable proxy function [1], [7], [8]. Similarly, entropy coding is invertible, but it is necessary to compute the length of the bitstream b.…”

Learned Lossless Image Compression with A Hyperprior and Discretized Gaussian Mixture Likelihoods

“…For example, some image compression approaches use generative models to learn the distribution of images using adversarial training [6,7,8] to achieved impressive subjective quality at extremely low bit rate. Some works use recurrent neural networks to compress the residual information recursively, such as [9,10,11] to realize scalable coding. Some approaches propose a hyperpriorbased and context-adaptive context model to compress codes effectively in [12,13,14].…”