DINER: Disorder-Invariant Implicit Neural Representation
This addresses a fundamental bottleneck in INR for inverse problems, offering improved performance in various domains, though it appears incremental as it builds on existing INR backbones.
The paper tackles the spectral bias limitation in implicit neural representations (INR) by proposing DINER, which uses a hash-table to re-arrange input coordinates, leading to significantly alleviated spectral bias and showing superiority over state-of-the-art methods in quality and speed across tasks like image/video representation and phase retrieval.
Implicit neural representation (INR) characterizes the attributes of a signal as a function of corresponding coordinates which emerges as a sharp weapon for solving inverse problems. However, the capacity of INR is limited by the spectral bias in the network training. In this paper, we find that such a frequency-related problem could be largely solved by re-arranging the coordinates of the input signal, for which we propose the disorder-invariant implicit neural representation (DINER) by augmenting a hash-table to a traditional INR backbone. Given discrete signals sharing the same histogram of attributes and different arrangement orders, the hash-table could project the coordinates into the same distribution for which the mapped signal can be better modeled using the subsequent INR network, leading to significantly alleviated spectral bias. Experiments not only reveal the generalization of the DINER for different INR backbones (MLP vs. SIREN) and various tasks (image/video representation, phase retrieval, and refractive index recovery) but also show the superiority over the state-of-the-art algorithms both in quality and speed.