EGD-Net: Eigenimage Guided Diffusion Network for Hyperspectral Mixed Noise Removal

In this paper, we study diffusion-type network methods for denoising remote sensing images with hyperspectral mixed noise (Gaussian noise and stripe noise). Two key issues should be addressed: (i) there are many wavelengths in remote sensing images, so it is necessary to reduce the computational work in diffusion networks for such high-dimensional data involved in the training procedure; (ii) as both Gaussian noise and stripe noise are considered, it is required to develop non- Gaussian-type diffusion networks. The main contribution of this paper is to design EGD-Net: Eigenimage Guided Diffusion Network for hyperspectral mixed noise removal. To address the raised issues, we propose to use a subspace representation in diffusion network so that several eigenimages are enough for learning. The clean eigenimage within the diffusion model guided by the dominant eigenimage undergoes gradual contamination by the corresponding noisy eigenimages over time instead of pure Gaussian noise. Therefore, the proposed guided diffusion model is not limited to deal Gaussian noise only. We evaluate the performance of EGD-Net on simulated hyperspectral images corrupted by Gaussian noise or mixed noise, as well as on real images. Our experimental results show that the performance of the EGD-Net is better than that of state-of-the-art methods.