FIPER: Generalizable Factorized Fields for Joint Image Compression and Super-Resolution

• DOI: 10.48550/arxiv.2410.18083
• Publication Date: 2024-10-23
• Authors: Yinghao Sun; Cheng Yu Yeo; Ernie Chu; Juncheng Chen; Yu-Lun Liu
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV); FOS: Electrical engineering, electronic engineering, information engineering

In this work, we propose a unified representation for Super-Resolution (SR) and Image Compression, termed **Factorized Fields**, motivated by the shared principles between these two tasks. Both SISR Compression require recovering preserving fine image details--whether enhancing resolution or reconstructing compressed data. Unlike previous methods that mainly focus on network architecture, our proposed approach utilizes basis-coefficient decomposition to explicitly capture multi-scale visual features structural components in images, addressing core challenges of both We first derive SR model, which includes Coefficient Backbone Basis Swin Transformer generalizable Factorized Fields. Then, further unify tasks, leverage strong information-recovery capabilities trained modules as priors compression pipeline, improving efficiency detail reconstruction. Additionally, introduce merged-basis branch consolidates structures, optimizing process. Extensive experiments show delivers state-of-the-art performance, achieving an average relative improvement 204.4% PSNR over baseline 9.35% BD-rate reduction compared SOTA.