Variational Autoencoder for Low Bit-rate Image Compression

We present an end-to-end trainable image compression framework for low bit-rate image compression. Our method is based on variational autoencoder, which consists of a nonlinear encoder transformation, a uniform quantizer, a nonlinear decoder transformation and a post-processing module. The prior probability of compressed representation is modeled by a Laplacian distribution using a hyperprior autoencoder and it is trained jointly with the transformation autoencoder. In order to remove the compression artifacts and blurs for low bit-rate images, an effective convolution based post-processing module is proposed. Finally, a rate control algorithm is applied to allocate the bits adaptively for each image, considering the bits constraint of the challenge. Across the experimental results on validation and test sets, the optimized framework trained by perceptual loss generates the best performance in terms of MS-SSIM. The results also indicate that the proposed postprocessing module can improve compression performance for both deep learning based and traditional methods, with the highest PSNR as 32.09 at the bit-rate of 0.15.