FLAIR: Frequency- and Locality-Aware Implicit Neural Representations

2025-08-19

AI Summary

Implicit Neural Representations (INRs) suffer from spectral bias, hindering faithful modeling of high-frequency details due to inherent deficiencies in frequency selectivity, spatial locality, and sparse representation capability. To address this, we propose FLAIR—a novel INR framework featuring the RC-GAUSS activation function, which jointly enables frequency-domain selection and spatial localization, and a Wavelet Energy-Guided Encoding (WEGE) mechanism that dynamically allocates frequency bandwidth via discrete wavelet transform (DWT), thereby circumventing the time-frequency uncertainty principle. FLAIR preserves the standard coordinate-to-signal mapping architecture while significantly enhancing representational sparsity and high-frequency fidelity. Extensive experiments demonstrate that FLAIR consistently outperforms state-of-the-art INRs on both 2D image reconstruction and 3D scene recovery tasks, effectively mitigating spectral bias and substantially improving fine-detail reconstruction accuracy.

Technology Category

• Computer Vision: Representation Learning for Vision
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Machine Learning: Representation Learning

Application Category

• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Implicit Neural Representations (INRs) leverage neural networks to map coordinates to corresponding signals, enabling continuous and compact representations. This paradigm has driven significant advances in various vision tasks. However, existing INRs lack frequency selectivity, spatial localization, and sparse representations, leading to an over-reliance on redundant signal components. Consequently, they exhibit spectral bias, tending to learn low-frequency components early while struggling to capture fine high-frequency details. To address these issues, we propose FLAIR (Frequency- and Locality-Aware Implicit Neural Representations), which incorporates two key innovations. The first is RC-GAUSS, a novel activation designed for explicit frequency selection and spatial localization under the constraints of the time-frequency uncertainty principle (TFUP). The second is Wavelet-Energy-Guided Encoding (WEGE), which leverages the discrete wavelet transform (DWT) to compute energy scores and explicitly guide frequency information to the network. Our method consistently outperforms existing INRs in 2D image representation and restoration, as well as 3D reconstruction.

Problem

• Addresses lack of frequency selectivity in neural representations
• Solves spatial localization issues in implicit neural networks
• Overcomes spectral bias in learning high-frequency details

Innovation

• RC-GAUSS activation for frequency selection
• WEGE encoding using wavelet energy
• Addresses spectral bias in neural representations