Implicit neural representation based on adaptive learnable filters
Yuhang Zheng, Ligen Shi, Chang Liu, Jun Qiu · 2024
Implicit Neural Representation (INR) achieves the mapping between coordinates and signal values through multilayer perceptrons (MLPs). However, coordinate-based MLPs struggle with high-frequency information due to spectral bias. A common solution to the spectral bias problem is to use Fourier features with fixed encoding frequencies instead of spatial coordinates as inputs to the MLPs. Natural scenes exhibit distinct frequency spectra across local areas, with the main regions characterized by low frequencies and the edges by high frequencies. The fixed encoding frequency Fourier feature will lead to redundant input information and reduce the running speed of MLPs. Therefore, this paper proposes an Implicit Neural Representation with Adaptive Learnable Filters framework (NRALF), which is controlled by a learnable parameter to achieve low-pass, band-pass, high-pass, all-pass and all-no-pass functions. This filter filters the Fourier features of different frequencies, making the filtered coding space more sparse, thereby enhancing the expression ability of MLPs for natural scenes and significantly improving the convergence speed of MLPs. Experimental results show that the proposed method has higher accuracy and faster convergence speed in scene representation.