Hyperspectral image denoising based on multi-resolution dense memory network

Hyperspectral images (HSIs) denoising is an important pre-processing step since noise will seriously degrade the HSIs quality. In this paper, we propose a multi-resolution dense memory denoising network. Specifically, in order to fully use the spatial-spectral correlation of hyperspectral image (HSI), target noisy band and multiple adjacent bands of HSI are extracted as our network input. After feature extraction by convolution kernels, the feature map is divided into resolution of different sizes through average pooling operation. For each resolution, according to the power law distribution of memory system, we design a dense connection structure in which we use weighted sum of the output feature maps for all previous layers. Then, the deconvolution is used for up sampling. Finally, dilated convolution and skip connection are utilized to obtain the denoised HSI. Our method outperforms state-of-the-art denoising algorithms in both quantitative metrics (e.g. PSNR and SSIM) and visual effects on simulated datasets. The real datasets (Indian Pines) experiment demonstrates that the classification accuracy of the denoised HSI is improved by 20.26% compared with the noisy HSI.