A Generalizable Model for Hyperspectral Image Denoising Based on Implicit Neural Representation

As multiple types of noise inevitably degrade the quality of hyperspectral images, denoising is a crucial preprocessing step for subsequent applications. Although many current deep learning-based methods have demonstrated satisfactory performance in hyperspectral image denoising, they often assume that the real data have the same noise distribution as the simulated training data. However, due to instrumental factors and the complex imaging environment, noise distribution in real-world scenarios exhibits diverse distributions. Consequently, their denoising performance usually drops significantly when applied to out-of-distribution noisy images. In this paper, we propose a generalizable hyperspectral image denoising method based on implicit neural representation (INR), denoted as Deinr, which can achieve competitive denoising results within known noise distributions while maintaining robust performance in out-of-distribution cases. Specifically, we construct a convolutional neural network to extract spatial-spectral features of observed HSIs. Following that, these features are fused with position vectors and then fed into several multi-layer perceptron (MLP) layers for coordinate-to-pixel mapping by the learnable implicit function. To capture high-frequency information effectively, the position vectors are derived through position encoding on coordinates, thereby enhancing the representation capability of the network. Moreover, additional MLP layers are applied to modulate the coordinate-to-pixel mapping features to enhance the generalization ability of the INR model. Extensive experiments conducted on simulated and real-world datasets demonstrate that the proposed method outperforms other state-of-the-art denoising methods in both in-distribution and out-of-distribution cases.