Compression and noise reduction of hyperspectral images using non-negative tensor decomposition and compressed sensing

Abstract

Hyperspectral images (HSI) are usually volumetric and require alot of space and time for archiving and transmitting. In this research, a new lossy compression method for HSI is introduced based on non-negative Tucker decomposition (NTD). This method consider HSI as a 3D dataset: two spatial dimensions and one spectral dimension. The NTD algorithm decomposes the original data into a smaller 3D dataset (core tensor) and three matrices. In the proposed method, the Block Coordinate Descent (BCD) method is used to find the optimal decomposition, which is initialized by using Compressed Sensing (CS). The obtained optimal core tensor and matrices are coded by applying arithmetic coding and finally the compressed dataset is transmitted. The proposed method is applied to the real dataset, simulation results show that in comparison with well-known lossy compression methods such as 3D SPECK and PCA+JPEG2000, the proposed method achieves the highest signal to noise ratio (SNR) at any desired compression ratio (CR) while noise reduction is simultaneously acquired.