Compressive sensing reconstruction of hyperspectral images based on codec space‐spectrum joint dense residual network. Issue 3 (24th November 2022)

• Record Type: Journal Article
• Authors: Xiao, Shuming; Zhang, Ye; Chang, Xuling; Xu, Jiajia
• Is Part Of: IET image processing. Volume 17:Issue 3(2023)
• Page Start: 916
• Page End: 931
• Publication Date: 2022-11-24
• Subjects: Image processing -- Periodicals; 621.36705
• DOI: 10.1049/ipr2.12682
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract: The spatial and spectral information contained in the hyperspectral image (HSI) make it widely used in many fields. However, the sharp increase of HSI data brings enormous pressure to the data storage and real‐time transmission. The research shows that hyperspectral compressive sensing (HCS) breaks through the bottleneck of the Nyquist sampling theorem, which can relieve the massive pressure on data storage and real‐time transmission. Existing HCS methods try to design advanced compression sampling matrix or reconstruction algorithms, but cannot connect the two through a unified framework. To further improve the image reconstruction quality, a novel codec space‐spectrum joint dense residual network (CDS2‐DResN) is proposed. The CDS2‐DResN is divided into block compression sampling part and reconstruction part. For block compression sampling, coded convolutional layer (CCL) is leveraged to compress and sample HSI. For measurements reconstruction, deconvolution layer is first leveraged to initially reconstruct HSI, and then build a space‐spectrum joint network to refine the initial reconstructed HSI. Moreover, the CCL and reconstruction network are optimized via a unified framework, which can simplify the pre‐processing and post‐processing process of HCS. Extensive experiments have shown that CDS2‐DResN has an excellent HCS reconstruction effect at measurement rates 0.25, 0.10, 0.04 and 0.01, respectively. Abstract : To further enhance the quality of image reconstruction in HCS, we propose a novel codec space‐spectrum joint dense residual network (CDS2‐DResN). The CDS2‐DResN jointly trains the block compression sampling phase and compressive sensing (CS) measurements reconstruction phase. By leveraging the self‐learning ability of convolutional neural networks (CNN), the CCL and the reconstruction network are connected and optimized via a unified framework.