研究目的
To propose a new scheme for the compression and reconstruction of hyperspectral images acquired in a compressive sensing fashion, aiming to reduce computational resources and improve reconstruction accuracy compared to existing techniques.
研究成果
The proposed SHSIR algorithm effectively compresses and reconstructs hyperspectral images by leveraging high spatial correlation and a small number of endmembers, using Bregman iterations for improved accuracy. Experimental results show superior performance in SSIM, PSNR, and SAM compared to state-of-the-art methods, demonstrating its effectiveness for hyperspectral data processing in satellite communications.
研究不足
The main constraint of the SHSIR algorithm is computation time due to the large size of hyperspectral datasets, but this is overlooked as reconstruction occurs at the ground station where resources are abundant. The method assumes prior estimation of the mixing matrix using RMVSA, which may not always be practical.
1：Experimental Design and Method Selection:
The methodology involves generating compressed measurements using Gaussian i.i.d. measurement matrices, formulating the reconstruction as a constrained optimization problem with non-smooth terms, and solving it using an adaptive Bregman iterations method of multipliers to convert it into a cyclic sequence problem.
2：Sample Selection and Data Sources:
Two hyperspectral datasets are used: URBAN (dimensions 307×307×210, cropped to 200×200 pixels per band) and PAVIAU (dimensions 610×340×103, cropped to 200×200 pixels per band).
3：List of Experimental Equipment and Materials:
A computer system with 8GB RAM, 1TB ROM, Intel i5 processor, 2GB NVidia Graphics card, Windows 10 OS, and MATLAB 2016b software.
4：Experimental Procedures and Operational Workflow:
The SHSIR algorithm is implemented, which includes steps for initialization, parameter selection, updates, and Bregman iterations until a stopping criterion is met. Performance is evaluated using SSIM, PSNR, and SAM metrics at sampling rates from
5：1 to 5, with added Gaussian noise to simulate SNR of 20 dB. Data Analysis Methods:
The reconstructed images are compared with existing methods (OMP, RLPHCS, SSHBCS) using structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and spectral angle mapper (SAM) to assess performance.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
