ours data

Implicit Neural Representations (INR) utilize Multi-Layer Perceptrons (MLPs) to map coordinates into signal values and have been widely applied in scene modeling. MLPs exhibit a strong preference for low-frequency components, resulting in poor performance when capturing high-frequency details. Early methods enhanced the high-frequency representation capability of MLPs by introducing Fourier feature maps. However, the encoding dimension requires careful tuning to capture high-frequency information while avoiding unnecessary high-frequency oscillations. In response to these issues, we propose an implicit neural representation method combining MLPs with Kolmogorov\u2013Arnold Networks (KANs) featuring learnable activation functions, enabling independent processing of low- and high-frequency information in signals. MLPs learn low-frequency components, while KANs with learnable activation functions capture high-frequency information. To further enhance the model's representational power across different frequency components, we incorporate two Discrete Wavelet Transform (DWT) and one Inverse Discrete Wavelet Transform (IDWT) into the network architecture. Forward decomposition splits the low- and high-frequency signals learned by MLPs and KANs into multiple frequency subbands, while backward reconstruction integrates these subbands to achieve high-quality signal reconstruction. Experimental results demonstrate that the proposed method outperforms current mainstream INR methods in scene representation accuracy. When processing natural scenes with complex frequency distributions, this method exhibits superior robustness and detail preservation capabilities.