Video compressive sensing with redundant dictionary

Compressive sensing is an innovative theory which allows us to sample signals under random projection domain. This technique seeks to minimize the cost of redundant data acquisition. In this paper, we propose a new video acquisition system which samples the video volumes with far fewer measurements than traditional camera. Video is divided into little time-spatial volumes due to diverse scene content change among frame regions. With strict sparsity constraints, adaptive dictionary is trained to obtain best representation for little video volumes. In this scheme, K-means clustering and KSVD learning are applied to selected video patches. Experiments and simulation are conducted to test the performance of the capability and adaptivity of the dictionary. Also, visual and PSNR comparison for video acquisition are provided to demonstrate the power of our system. We show that our approach can effectively reconstruct the original video with as few as 5% measurements without losing spatial or temporal resolution.