NIF25: Reducing the Gap in INR-Based Image Compression

Le résumé fourni par la source

Implicit Neural Representations (INRs) have recently proven to be effective in data compression tasks, offering an alternative to complex hand-crafted encoding and decoding pipelines. However, INR-based methods for image compression suffer from long encoding times and a consistent performance gap in classic quality metrics such as PSNR. In this work, we present NIF25, a novel image compression codec based on Implicit Neural Representations, which advances the current state-of-the-art NIF codec by consistently improving the rate-distortion ratio while maintaining comparable compression and decompression speed. We evaluate our propoposal against INR-based, traditional and hybrid image codecs considering quantitative metrics, computational times, and perceived visual fidelity. Our results show that NIF25 enhances the INR-based compression performance by approximately 25% while reducing information loss. In addition, this novel codec excels in modern metrics such as LPIPS, outperforming or matching well-established and contemporary methods by offering superior visual fidelity even at very low bitrates. An extensive ablation study highlights the contribution of each component of the proposed design, leading the way for further developments and a deep understanding of the properties and capabilities of INR-based methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.