Streaming Neural Images

TL;DR

Implicit Neural Representations offer a continuous, resolution-agnostic way to encode images but face practical bottlenecks in compression due to cost, instability, and robustness. This work analyzes these limitations across INR families (Fourier Feature Networks, SIREN, MFN, DINER) and introduces SPINR, a Streaming Progressive INR that trains in stages and transmits the network progressively with layer-wise redundancy. Experimental results on Set5/Set14 show small, low-complexity INRs can achieve competitive rate-distortion, while larger models underperform; SPINR improves robustness to packet loss and enables adaptive bitrate through staged decoding. The findings provide a nuanced baseline for implicit neural image compression and guide future research toward robust, streaming INR codecs.

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.

Figures (8)

• Figure 1: Exploring the behaviour of neural image representations. (Left) An image after losing one random pixel. (Mid) The corresponding implicit neural representation (INR) sitzmann2020implicit. (Right) The INR network after losing one random neuron.
• Figure 2: We illustrate the general concepts around neural image representations tancik2020fouriersitzmann2020implicit. We also illustrate the common frameworks for streaming images as INRs dupont2022coin++strumpler2022implicit. This can be extended to other sort of signals such as audio or 3D representations.
• Figure 3: Image streaming using (top) traditional image representations and codecs pennebaker1992jpeg, (bot.) our method, SPINR, based on implicit neural image compression sitzmann2020implicitdupont2021coin allows to decode the image without having the full neural network.
• Figure 4: Comparison between INR compression strumpler2022implicitdupont2021coindupont2022coin++ and classical JPEG compression using the Kodak dataset. Best viewed in electronic version.
• Figure 5: Training evolution of the different INR methods. We observe high training instability for DINER xie2023diner. The models have $h=128, l=4$. We also show the corresponding image.