Spatially adaptive image compression using a tiled deep network.

Deep neural networks represent a powerful class of function approximatorsthat can learn to compress and reconstruct images. Existing image compressionalgorithms based on neural networks learn quantized representations with aconstant spatial bit rate across each image. While entropy coding introducessome spatial variation, traditional codecs have benefited significantly byexplicitly adapting the bit rate based on local image complexity and visualsaliency. This paper introduces an algorithm that combines deep neural networkswith quality-sensitive bit rate adaptation using a tiled network. Wedemonstrate the importance of spatial context prediction and show improvedquantitative (PSNR) and qualitative (subjective rater assessment) resultscompared to a non-adaptive baseline and a recently published image compressionmodel based on fully-convolutional neural networks.