Streaming Neural Images

TL;DR

This paper investigates the limitations of implicit neural representations for image compression, focusing on computational cost, stability, and robustness, and provides empirical insights into methods like Fourier Feature Networks and Siren.

Contribution

It offers a comprehensive analysis of the challenges in INR-based image compression and highlights areas for future research to improve performance and reliability.

Findings

INRs face significant computational and stability challenges.

Empirical analysis of Fourier Feature Networks and Siren reveals their strengths and weaknesses.

Insights provided can guide future improvements in neural image compression.

Abstract

Implicit Neural Representations (INRs) are a novel paradigm for signal representation that have attracted considerable interest for image compression. INRs offer unprecedented advantages in signal resolution and memory efficiency, enabling new possibilities for compression techniques. However, the existing limitations of INRs for image compression have not been sufficiently addressed in the literature. In this work, we explore the critical yet overlooked limiting factors of INRs, such as computational cost, unstable performance, and robustness. Through extensive experiments and empirical analysis, we provide a deeper and more nuanced understanding of implicit neural image compression methods such as Fourier Feature Networks and Siren. Our work also offers valuable insights for future research in this area.