Few Pains, Many Gains: Fast On-device Image Compression through Super Resolution

Mobile terminals have been accepted as main stream tools for online sociality, from which people upload photos and videos to share their lives. However, the huge expense of intensive bandwidth occupation boosts the running cost of service vendors. Conventional image compression methods alleviate this problem but fail to guarantee the visual quality and operation efficiency simultaneously. Therefore, in this paper, we propose a lightweight encoder that introduces super resolution methods to save bandwidth. In the pipeline, images are downsized at sources, transmitted with less bandwidth requirements, and reconstructed at sinks. The downscaling operation is computational friendly for most mobile terminal processors while the complex reconstructing calculation is assigned to powerful servers of the sharing platform. To accelerate the process and improve reconstructing quality, we propose an edge feature fusion super resolution method embedded in an efficient deep neural network for high- resolution image upload. Our method eliminates the heavy computation brought by the encoding process while guaranteeing reconstruction quality. Quantitative and qualitative evaluations on benchmarks demonstrate that in the same encoding approach, our performance surpass the state-of-the-art while largely reducing parameters and computations.