基于颜色自适应字典的双相机光谱系统重建算法

Compared with the conventional RGB image and panchromatic image, hyperspectral image can provide more details and features with the additional spectral dimension. Nowadays, hyperspectral image has been applied into various computer vision tasks, such as classification, medical diagnosis, face recognition, objects tracking, and so on. In order to obtain hyperspectral image, traditional imaging systems capture the 3D information with scanning techniques. Such scanning based imaging systems can only record the spectral information of one or a few scene points at the same time, which inevitably suffers from the tradeoff between spectral resolution and time effciency. Thanks to the flourish of computational photography, snapshot spectral imagers have been developed to overcome the drawback of conventional imaging systems in recent years. Imaging systems in this category have the ability to capture the full hyperspectral image with one single exposure. Leveraging the compressive sensing (CS) theory, coded aperture snapshot spectral imaging (CASSI) stands out as a promising solution among those systems. With elaborate optical design, CASSI encodes the 3D hyperspectral image into the 2D compressive measurement and then reconstruct the underlying image with the CS theory. Incorporating a CASSI system and a color detector, the dual-camera compressive hyperspectral imager can promote the reconstruction accuracy of CASSI efficiently and thus owns broad application prospects. How to reconstruct hyperspectral image from the compressive measurement with high quality is an urgent problem to be solved for the system. The existing methods exploit a single prior information of hyperspectral image to develop reconstruction algorithms, which fail to make full use of the measurement of dual-camera design, and the reconstruction quality is not ideal enough. Based on the strong correlation between hyperspectral image and its corresponding color image in spatial structure and spectral response, we propose a color adaptive dictionary based reconstruction method to improve the reconstruction quality. First, in the case of introducing non-negative constraints to dictionary elements and sparse representation coefficients, three over-complete dictionaries are learned from the color measurement. Considering the fact that a single band of hyperspectral image owns very high texture and structure similarity with its corresponding color measurement, we utilize the color measurement to learn over-complete dictionaries for sparse reconstruction. Second, based on the spectral response of the color camera, a suitable dictionary with high spectral correlation is selected for each band. Specifically, we choose the dictionary of the channel with the largest response amplitude of the RGB camera in the current spectral band as its sparse basis, so as to ensure the high sparsity of the sparse representation and improve the reconstruction quality. Then, by integrating the sparse representation with the system imaging principle, we develop an optimization framework for hyperspectral image reconstruction, which is finally solved via the alternative direction multiplier method. At last, we conduct a thorough experiment on both the hyperspectral and the remote sensing data sets to validate the performance of our method. Simulation results suggest that the proposed method can greatly improve the reconstruction fidelity of the dual-camera compressive hyperspectral imager, which verifies the practical application potential of our method.