Learning Content-Weighted Deep Image Compression

Learning-based lossy image compression usually involves the joint optimization of rate-distortion performance, and requires to cope with spatial variation content contextual dependence among learned codes. Traditional entropy models can spatially adapt local bit rate based on content, but are limited in exploiting context code space. On other hand, most deep computationally very expensive cannot efficiently perform decoding over symbols parallel. In this paper, we present a content-weighted encoder-decoder model, where channel-wise multi-valued quantization is deployed for discretization encoder features, an importance map subnet introduced generate masks varying pruning. Consequently, summation serve as upper bound length bitstream. Furthermore, quantized representations still dependent, which be losslessly compressed using arithmetic coding. To compress codes effectively efficiently, propose upper-triangular masked convolutional network (triuMCN) large modeling. Experiments show that proposed method produce visually much better results, performs favorably against traditional approaches.