TLIC: Learned Image Compression with ROI-Weighted Distortion and Bit Allocation

TL;DR

The paper targets perceptual quality in learned image compression by combining ROI-weighted distortion and ROI-guided bitrate allocation with adversarial texture synthesis. ROI maps from RMformer generate $Mask_{2D}$ and smoothed ROI maps that steer per-pixel distortion and channel-wise bit allocation via a weighted encoder, with a two-stage training regime that first optimizes MSE and then perceptual quality using a U‑Net discriminator and auxiliary losses. A gain-based, continuous-rate control framework and a simplified entropy model enable flexible target bitrate while preserving important regions. The approach demonstrates improved perceptual realism at low bitrates by protecting background content and guiding allocation with ROI information.

Abstract

This short paper describes our method for the track of image compression. To achieve better perceptual quality, we use the adversarial loss to generate realistic textures, use region of interest (ROI) mask to guide the bit allocation for different regions. Our Team name is TLIC.