PO-RTIntra: Perception-Weighted Rate Allocation and ROI Latent Optimization on DCVC-RT Intra

Wenzhuo Ma, Junxi Zhang, Nianxiang Fu, Zhenzhong Chen

Abstract: In this paper, we propose PO-RTIntra, a perception-oriented learned image compression framework built upon the DCVC-RT intra model, together with a higher-capacity variant, PO-RTIntraPro. The strong compression performance and efficiency of the DCVC-RT intra model provide a solid backbone for PO-RTIntra, while PO-RTIntraPro increases the capacity of key modules to further enhance modeling capacity. We adopt a multi-stage progressive training schedule and incorporate a composite perceptual loss together with a Relativistic PatchGAN discriminator to improve perceptual fidelity. In addition, we introduce a Human-Perception-weighted Integer Linear Programming (HP-ILP) formulation for bitrate allocation, and an ROI-based Latent Rate–Distortion-Optimized (ROI-LRDO) inference strategy to further improve reconstruction quality. Experiments demonstrate that, compared with state-of-the-art image compression methods, our approach produces more realistic, detail-rich reconstructions.

Team Name: Thanos, IronMan

Submission Number: 8