I-INR: Iterative Implicit Neural Representations

TL;DR

This work addresses the limited high-frequency fidelity and noise robustness of implicit neural representations by introducing Iterative Implicit Neural Representations (I-INR), a plug-and-play framework that refines INR outputs over multiple steps using a BackboneNet complemented by lightweight FeedbackNet and FuseNet. By sharing the Backbone computation and applying iterative refinement, I-INR achieves superior detail preservation and robustness across image fitting, super-resolution, denoising, and 3D occupancy tasks with minimal overhead. The approach is validated through extensive experiments showing consistent improvements over SIREN, WIRE, and Gauss in both quantitative metrics (PSNR, SSIM, LPIPS) and qualitative reconstructions. The results suggest a practical pathway to deploy more faithful INRs in real-world noisy and incomplete data scenarios.

Abstract

Implicit Neural Representations (INRs) have revolutionized signal processing and computer vision by modeling signals as continuous, differentiable functions parameterized by neural networks. However, their inherent formulation as a regression problem makes them prone to regression to the mean, limiting their ability to capture fine details, retain high-frequency information, and handle noise effectively. To address these challenges, we propose Iterative Implicit Neural Representations (I-INRs) a novel plug-and-play framework that enhances signal reconstruction through an iterative refinement process. I-INRs effectively recover high-frequency details, improve robustness to noise, and achieve superior reconstruction quality. Our framework seamlessly integrates with existing INR architectures, delivering substantial performance gains across various tasks. Extensive experiments show that I-INRs outperform baseline methods, including WIRE, SIREN, and Gauss, in diverse computer vision applications such as image restoration, image denoising, and object occupancy prediction.

Figures (15)

• Figure 1: Effectiveness of the proposed methods across multiple tasks, including image fitting, super resolution (2$\times$ scale), and denoising, compared to the baseline representative INR method (SIREN). Our novel Iterative-INR method (I-SIREN) consistently improves detail preservation, fidelity, and high-frequency reconstruction across all tasks, outperforming the baseline.
• Figure 2: (a) The proposed architecture of the I-INR model. The framework consists of a Backbone, which can be any baseline INR architecture. Additionally, a Feedback Net incorporates feedback to refine representations, while FuseNet integrates features. The final output is obtained by combining the outputs of the Backbone and the FuseNet, enhancing expressivity and reconstruction quality. (b) Iterative reconstruction process of the proposed I-INR framework.
• Figure 3: Image fitting results for different non-linearities and their iterative extensions (prefix "I"). Each column presents the reconstructed image (top) and residual map (bottom), highlighting reconstruction quality. PSNR values are reported, with iterative improvements in red. The rightmost column shows the ground truth.
• Figure 4: Visual quality comparison of super-resolution results at 2$\times$ scale using various methods. The ground truth is compared against baseline INR methods SIREN, WIRE, Gauss, and their iterative counterparts. The iterative approaches consistently achieve sharper reconstructions with fewer artifacts, effectively preserving finer details and high-frequency structures.
• Figure 5: Visual comparison of denoising results using various methods. The ground truth is compared against noisy input, baseline methods SIREN, WIRE, and Gauss, as well as their iterative counterparts. I-INR demonstrates superior artifact reduction and detail preservation compared to their non-iterative counterparts.