Deep Tensor Attention Prior Network for Hyperspectral Image Denoising

Hyperspectral imaging techniques can generate continuous narrowband images with a high spectral resolution. However, owing to environmental disturbances, atmospheric effects, and hardware limitations of hyperspectral imaging sensors, captured hyperspectral images (HSIs) often contain complex noises that significantly degrades their quality and limits their utility. In this article, we design a deep tensor attention module based on a canonical-polyadic (CP) decomposition of a feature tensor, referred to as a deep CP attention module (DCPAM), which can disentangle spatial and channel information and further enhance the topological structure of features. It has been shown that the DCPAM is effective and relatively simple to integrate into some well-known network architectures. And taking the network with the DCPAM as a deep prior, we propose a deep tensor attention prior network (DTAPNet) for HSI denoising tasks. Extensive experiments on simulated and real HSIs demonstrate that our proposed DTAPNet outperforms existing state-of-the-art HSI denoising models.