Abstract
Implicit Neural Representations (INRs) have emerged as a powerful paradigm for representing continuous signals through coordinate-based neural networks. However, existing methods often struggle to balance high-frequency detail with spatial localization. This paper introduces TRIDENT, a novel INR function that unifies three core nonlinear properties: order, frequency, and spatial compactness. By mathematically integrating these pillars, TRIDENT achieves state-of-the-art performance across diverse modalities, including 3D reconstruction, audio recovery, and medical imaging, particularly excelling in underconstrained inverse problems.
Introduction
Implicit Neural Representations (INRs) represent signals—such as images, 3D shapes, or audio—as continuous functions mapped by neural networks. While transformative, current INR techniques face a persistent "Goldilocks" challenge: models either over-smooth data, losing critical fine-grained details, or introduce unwanted artifacts (like "ringing") when attempting to capture high-frequency information. A primary cause is the lack of "spatial compactness"—the ability to localize signal energy effectively without it bleeding into adjacent regions.
The TRIDENT Framework
The TRIDENT function is built upon a "trilogy" of theoretical pillars designed to overcome the limitations of previous Fourier-based and periodic activation functions:
- Order Compactness: The network represents high-order features more efficiently by elevating the series index to the exponent of the cosine function within a Fourier series.
- Frequency Compactness: Utilizing multi-scale sinusoidal functions inspired by Fourier Transformation, the model captures a wide spectrum of frequency information.
- Spatial Compactness: By incorporating a Gaussian-like windowing effect, TRIDENT ensures signal energy is well-localized, preventing the artifacts common in pure Fourier methods.
The synergy of these components is critical. As shown in the ablation study below, removing any single element of the trilogy leads to a measurable drop in reconstruction quality (SSIM).
Experimental Evaluation
TRIDENT was rigorously tested across multiple benchmarks to evaluate its versatility and robustness:
- Image Processing: Denoising on the DIV2K dataset and Super-Resolution on Set5/Set14.
- 3D Reconstruction: Modeling the complex geometry of the "Thai Statue" at high resolution.
- Audio Recovery: Reconstructing complex signals including human speech and classical music.
- Medical Imaging: Sparse-view CT reconstruction using chest images from the Cancer Imaging Archive (TCIA).
Key Results
TRIDENT consistently outperformed established methods like SIREN and WIRE:
- Denoising & Super-Resolution: Achieved a PSNR of 31.26 on the "Parrot" image (vs. WIRE's 30.05). It produced high-fidelity results with minimal color striation artifacts, as seen in the natural image reconstruction below.
- 3D Modeling: Reached a 99.29% Intersection over Union (IoU) for the Thai Statue, capturing intricate surface details.
- Underconstrained CT Reconstruction: In scenarios with only 50 projections, TRIDENT maintained an SSIM of 0.78, significantly higher than WIRE (0.61), demonstrating exceptional stability with missing data.
Discussion and Limitations
The primary innovation of TRIDENT lies in its multimodal robustness. Unlike specialized INRs, it accurately reconstructs both the temporal nuances of audio and the spatial complexities of 3D volumes. However, the model's performance is sensitive to the $s_0$ (scale) parameter, which controls the spatial window. A value that is too small loses low-frequency data, while a value too large compromises spatial compactness. Currently, an optimal value (typically around 5) must be determined empirically.
Conclusion
TRIDENT provides a mathematically grounded solution to the trade-offs inherent in implicit neural representations. By unifying order, frequency, and spatial compactness, it offers a robust framework for high-fidelity signal representation across a wide array of scientific and creative applications. Future research will focus on automated hyperparameter tuning and extending the framework to 4D spatio-temporal data.
Summary updated