End-to-end Optimized Image Compression with Attention Mechanism

Lei Zhou, Zhenhong Sun, Xiangji Wu, Junmin Wu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

We present an end-to-end trainable image compression framework for low bit-rate and transparent image compression. Our method is based on variational autoencoder, which consists of a nonlinear encoder transformation, a soft quantizer, a nonlinear decoder transformation and a entropy estimation module. The prior probability of the latent representations is modeled by combining a hyperprior autoencoder and a Pixelcnn++ based context module and they are trained jointly with the transformation autoencoder with attention mechanism. In order to improve the compression performance, a non-local convolution based attention mechanism is designed for allocating bits adaptively. Finally, a novel rate allocation algorithm based on linear optimization is used to assign the bits for each image dynamically, considering the bits constraint of the challenge. Across the experimental results on validation and test sets, the optimized framework can generate the highest PSNR and MS-SSIM for low bit-rate compression competition, and cost the lowest bytes for transparent 40db competition.