SASNet: Spatially-Adaptive Sinusoidal Neural Networks
This addresses challenges in computer vision and graphics for researchers and practitioners using implicit neural representations, offering a robust method for high-frequency signal reconstruction, though it is incremental as it builds on existing sinusoidal neural networks.
The paper tackled the problem of spectral bias, training instability, and overfitting in sinusoidal neural networks for implicit neural representations, proposing SASNet which integrates frequency embedding and spatially-adaptive masks to enhance signal reconstruction, achieving strong fitting accuracy and super-resolution capability without sacrificing compactness.
Sinusoidal neural networks (SNNs) have emerged as powerful implicit neural representations (INRs) for low-dimensional signals in computer vision and graphics. They enable high-frequency signal reconstruction and smooth manifold modeling; however, they often suffer from spectral bias, training instability, and overfitting. To address these challenges, we propose SASNet, Spatially-Adaptive SNNs that robustly enhance the capacity of compact INRs to fit detailed signals. SASNet integrates a frequency embedding layer to control frequency components and mitigate spectral bias, along with jointly optimized, spatially-adaptive masks that localize neuron influence, reducing network redundancy and improving convergence stability. Robust to hyperparameter selection, SASNet faithfully reconstructs high-frequency signals without overfitting low-frequency regions. Our experiments show that SASNet outperforms state-of-the-art INRs, achieving strong fitting accuracy, super-resolution capability, and noise suppression, without sacrificing model compactness.