Attention-Based Deep Autoencoder for Hyperspectral Image Denoising

Hyperspectral Image (HSI) denoising is a crucial pre-processing task due to its widespread applications in areas that include geology, medicine, agriculture, surveillance, and the food industry. Denoising HSI data improves high-level vision tasks like classification, object tracking, and video surveillance. In this work, we propose a novel attention-based deep auto-encoder for HSI denoising which learns a mapping from noisy observation to clean the data. To exploit the spatial-spectral information present in HSI data, 3D symmetric convolution and deconvolution are used. The 3D features extracted from 3D convolutional layers (CL) helps the restoration across the depth dimension. However, feature maps obtained from the convolutional layers are localized and are not able to capture long-range dependencies in the data. Furthermore, CL does not differentiate between low and high degradation levels; producing smoothed results in low-textured areas of the image and artifacts in the high-textured regions. To avoid these adverse effects, attention blocks are plugged into the network to focus on more relevant features during the restoration process. Features obtained from attention blocks are fed as symmetric skip connections into the corresponding deconvolution layers. The proposed model is fully convolutional; it can process 3D images of arbitrary dimensions during training and inference. The experimental results show that our proposed model: Attention-based Deep Auto-encoder (AbDAE) outperforms the state-of-the-art methods in terms of visual as well as quantitative results.