An efficient and perceptiual-oriented image compression method

Jun Jiang

Abstract: The goal of our model is to compress a set of images to specified bitrates while preserving high perceptual quality, particularly at low bitrates. To this end, we employ a simple yet efficient mean-scale-hyperprior model, which effectively captures spatial dependencies in the latent representation through a hyperprior structure. We finetune the model using a perceptual quality-oriented loss function to enhance visual fidelity. To accommodate a range of target bitrates, we train four distinct models corresponding to different rate-distortion trade-off parameters, and dynamically select the most appropriate model for each input image based on its complexity and target bitrate.

Team Name: gauss

Submission Number: 5