I tried training a classifier, then found a better solution.
Abstract: Existing human visual perception-oriented image compression methods well maintain the perceptual quality of compressed images, but they may introduce fake details into the compressed images, ...