Adaptive Wavelet-Positional Encoding for High-Frequency Information Learning in Implicit Neural Representation

Hongxu Zhao, Zelin Gao, Yue Wang, Rong Xiong, Yu ZhangAAAI 2025 · DOI

TL;DR

Implicit Neural Representations (INRs) struggle to capture high-frequency details due to "spectral bias." This paper introduces Adaptive Wavelet-Positional Encoding (WPE) with High-Frequency Perception (HFP) to selectively embed high-frequency information, significantly improving INR performance in various tasks like 1D signal fitting, 2D image regression, and 3D scene representation.

What Problem Does This Paper Address?

Implicit Neural Representations (INRs) have emerged as a powerful paradigm for representing complex signals, such as 3D scenes and images, as continuous functions. Unlike traditional methods that rely on discrete representations, INRs learn a mapping from coordinates to signal values using neural networks (typically MLPs). However, a significant challenge, known as "spectral bias," plagues these methods. Neural networks inherently tend to learn low-frequency components of a signal more effectively than high-frequency ones. This leads to INRs producing blurry or incomplete representations, especially in areas with fine details or sharp transitions. Existing approaches to mitigate spectral bias, like standard Positional Encoding (PE) or Fourier Feature Networks (FFN), often apply global transformations, failing to address the non-uniform distribution of high-frequency information in real-world signals.

Methods

The paper proposes a two-pronged approach:

  • High-Frequency Perception (HFP): This novel method aims to identify where high-frequency components are located within a target signal. For image-based inputs, HFP uses a pyramidal Discrete Wavelet Transform (DWT) to decompose the image into different frequency bands. By analyzing the distribution of high-frequency components, HFP identifies "hotspots" in the signal space. For 3D scene representation, HFP leverages ray tracing to sample high-frequency rays and identify regions with rich details.
  • Adaptive Wavelet-Positional Encoding (WPE): Building on the insights from HFP, WPE selectively embeds coordinate inputs into a higher-dimensional wavelet-frequency space. It utilizes the local nature of wavelet transforms, specifically the Morlet wavelet, to process coordinates. The "hotspots" identified by HFP serve as centers for these local wavelet functions. This allows for more complex encoding in regions with high-frequency content and simpler encoding elsewhere. To ensure comprehensive representation, WPE also incorporates global Positional Encoding (PE) to capture low-frequency and global features. The final embedding is a concatenation of the local wavelet encoding and the global PE.

A key advantage of this approach is its modularity. WPE can be integrated into existing INR frameworks without modifying their core training procedures or network architectures.

Datasets

The paper evaluates its method on the following datasets:

  • 1D Signal Fitting: A "Girl With a Pearl Earring" image is used, with random row pixel values sampled as target signals.
  • 2D Image Regression: The "Girl With a Pearl Earring" image is used, with pixel-wise alternating patterns for training and testing.
  • 3D Scene Representation:
    • NeRF-Synthetic dataset: A standard benchmark for neural radiance fields.
  • Real-World LLFF dataset: A dataset of challenging real-world scenes.

No new benchmarks were introduced; the focus was on evaluating the proposed method on existing, well-established datasets.

Key Insights and Ideas

  • Spectral bias is spatially non-uniform: The paper highlights that high-frequency information isn't evenly distributed. Targeting these specific regions is more effective than global approaches.
  • Wavelets offer local perception: The inherent locality of wavelet transforms makes them ideal for capturing fine details in specific regions, complementing the global nature of traditional PE.
  • HFP is crucial for adaptive encoding: Identifying high-frequency regions before encoding is key to making the wavelet embedding adaptive and efficient.
  • Modularity is a strength: The ability to plug WPE into existing INR frameworks without retraining is a significant practical advantage.
  • Global + Local is better: Combining global PE with local wavelet encoding provides a more complete representation across all frequency bands.

Key Results

  • Quantitative Improvements: The proposed WPE method consistently outperforms global methods like PE, FFN, and WIRE across various metrics (PSNR, SSIM, LPIPS) in 1D signal fitting, 2D image regression, and 3D scene reconstruction tasks. For instance, in 3D NeRF-Synthetic experiments, the WPE method achieved the highest PSNR (28.81) and SSIM (0.94) compared to PE (28.07, 0.91), FFN (27.93, 0.91), and WIRE (28.49, 0.93).
  • Qualitative Improvements: The method excels at preserving and reconstructing fine details, as evidenced by visual comparisons in 2D and 3D tasks. Figures show sharper edges and more intricate textures in reconstructions using WPE.
  • Optimization Efficiency: While global methods might converge faster initially, WPE eventually surpasses them, achieving higher PSNR values, suggesting its ability to capture more complex details leads to better long-term performance.

Limitations and Future Work

  • Performance in low-frequency scenes: The paper notes that in scenes with inherently low-frequency information (e.g., "Horns"), the improvement from WPE is less significant, as global methods are already sufficient.
  • Convergence speed: The paper mentions that in the early stages of training, the network might take longer to converge due to the increased dimensionality of the embedding.
  • Future Work: The authors suggest exploring different wavelet families and further investigating the adaptive adjustment of wavelet parameters. They also mention conducting experiments on different representation frameworks to demonstrate broader applicability.

Summary updated

BibTeX
@inproceedings{zhao2025adaptive,
  author = {Hongxu Zhao and Zelin Gao and Yue Wang and Rong Xiong and Yu Zhang},
  title = {{Adaptive Wavelet-Positional Encoding for High-Frequency Information Learning in Implicit Neural Representation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i10.33132},
}
Figure