We propose a method for lossy image compression based on recurrent,convolutional neural networks that outperforms BPG (4:2:0 ), WebP, JPEG2000,and JPEG as measured by MS-SSIM. We introduce three improvements over previousresearch that lead to this state-of-the-art result. First, we show thattraining with a pixel-wise loss weighted by SSIM increases reconstructionquality according to several metrics. Second, we modify the recurrentarchitecture to improve spatial diffusion, which allows the network to moreeffectively capture and propagate image information through the network'shidden state. Finally, in addition to lossless entropy coding, we use aspatially adaptive bit allocation algorithm to more efficiently use the limitednumber of bits to encode visually complex image regions. We evaluate our methodon the Kodak and Tecnick image sets and compare against standard codecs as wellrecently published methods based on deep neural networks.
展开▼
