FCNR: Fast Compressive Neural Representation of Visualization Images

TL;DR

FCNR tackles the challenge of compressing tens of thousands of visualization images across viewpoints and timesteps by introducing a fast compressive neural representation that leverages stereo information. Building on ECSIC, FCNR introduces joint context transfer modules (JCTMs) and stereo context modules (SCMs) to exploit mutual information between paired views, while integrating visualization parameters via positional encoding to improve entropy estimation. The method uses a Laplacian-based entropy model and differentiable quantization within a rate-distortion framework, achieving significantly faster encoding/decoding than prior INR-based methods and competitive compression ratios, often outperforming ECSIC in both quality (PSNR/LPIPS) and speed. Overall, FCNR offers a practical, high-fidelity solution for large-scale visualization data, with future work aimed at further tailoring architectures to visualization tasks and reducing bitrate while expanding parameter integration.

Abstract

We present FCNR, a fast compressive neural representation for tens of thousands of visualization images under varying viewpoints and timesteps. The existing NeRVI solution, albeit enjoying a high compression ratio, incurs slow speeds in encoding and decoding. Built on the recent advances in stereo image compression, FCNR assimilates stereo context modules and joint context transfer modules to compress image pairs. Our solution significantly improves encoding and decoding speed while maintaining high reconstruction quality and satisfying compression ratio. To demonstrate its effectiveness, we compare FCNR with state-of-the-art neural compression methods, including E-NeRV, HNeRV, NeRVI, and ECSIC. The source code can be found at https://github.com/YunfeiLu0112/FCNR.

Figures (5)

• Figure 1: Overview of FCNR. The encoder ($E$) encodes $x_l$ and $x_r$ to bitstreams with hyper-encoder ($h_E$), quantization ($Q$), and arithmetic encoder ($AE$). The decoder ($D$) then reconstructs $\hat{x}_l$ and $\hat{x}_r$ through quantized latents $(\hat{y}_l$ and $\hat{y}_r)$ with arithmetic decoder ($AD$) and hyper-decoder ($h_D$).
• Figure 2: The detailed structure of each module.
• Figure 3: Decompressed IR and DVR images. The datasets are vortex, Tangaroa, and tornado, respectively.
• Figure 4: PSNR comparison of all methods on the vortex IR dataset. E-NeRV, HNeRV, and NeRVI were all trained for 200 epochs. Both ECSIC and FCNR were trained for 3 epochs.
• Figure 5: First row: decompressed images of FCNR and ECSIC with variations. Second row: zoom-ins for closer examination.