Computational Spectral Imaging With Added Panchromatic Imaging And Its Parallel Reconstruction Algorithm

The power spectral density of light has great significance to show the components of the scene, which promote the research and its applications in remote sensing, geographic information system, environment monitoring, targets reconnaissance, etc. However, the spectral imformation varies significantly with spatial and temporal factors. It's necessary to capture a spectral image quickly and form a spectral video which includes more information.Based on he theory of compressed sensing, Duke University proposed a new computational spectral imager called coded aperture snapshot spectral imagers(CASSI), which can capture the power spectral density of light. In general, the whole spectral imaging process is divided into two stage: the observation process and the data restoration process. In the first stage, the observation process mainly obtain the measurements by coding and sampling of the spectral scene. In the second stage, the data restoration process can be turned into an inverse problem based on CS theory; that is to say, the 3D spectral data cube can be restored by solving a CS inverse problem.There are some drawbacks in the observation process, such as the lose of spectral information and the low sampling rate. The reconstruction quality mostly relies on the amount of measurements, the less sparsed scene the more measurements. For complex scenes which can't be sparsely represented under a fixed base, the recovery accuracy is poor due to the lack of measurements. Based on this, some papers put forward specific solutions for better better reconstuction quality, such as coded aperture optimization, multiframe image estimation and double channel system. In this paper, we propose a new compressive coded aperture spectral imaging method with an added panchromatic camera. By the proposed method we can effectively obtain more measurements to improve the recovery quality of the complex scenes. Furthermore, panchromatic image has a better spatial representation which contribute to improving the reconstructing accuracy of spatial structure. In the optimization reconstruction of original spectral scenes, we fully exploit the structural similarity between panchromatic image and the spectral data, this provide a fresh perspective for building a new compressive spectral imaging system. Simulation result shows that the proposed framework significantly improves the performance of compressive spectral imaging.For the more, data reconstruction surfer from a extremely high complexity. This paper presents a GPU based parallel implementation of the spectral reconstruction algorithm. Experiment results show that this presented fast algorithm effectively saves the spectral image quality and achieves 100 x accelaration over a CPU based implementation. When the size of the spectral video is 300x300,the speedup of GPU over CPU is more than 50 times, while when the size is 600x600 or 1200x1200, more that 130 times speedup can be achieved. These results verified the effectiveness of our algorithm.