Object-aware Image Compression with Adversarial Learning

摘要

Image compression based on region of interest (ROI) can maintain high quality on specific areas and behaves well in terms of subjective quality. However, existing methods lead to a severe loss of background information as well as block effects when the bitrate goes down. In this paper, we take advantage of the generative models and develop the generative compression scheme for ROI-based image compression. The proposed method adopts different compression objectives for interested areas and background. For a target bitrate, our method tends to preserve more details for interested areas while generating the background using a perceptual metric. The generative model works on the whole image to erase the unnatural block effects. Experiments show that our method could produce visually pleasing images at very low bitrates and is superior to existing compression methods.