Video Super Resolution Technology Based On Compressive Sensing

High resolution image can be reconstructed from one or multi-frames of low resolution images by super resolution technology. This technology is widely used in network videos, digital videos, public safety and so on. Compressive sensing is an innovative signal theory which breaks through the Nyquist sampling theory’s limit. The main research content of this paper is to use compressive sensing theory to solve the super-resolution problem.The paper describes the research background and significance of super resolution technology and compressive sensing theory, and summarizes the research status of single frame, multi-frames and video super-resolution technology. Theoretical framework and the specific application of compressive sensing also are introduced in this paper.A single frame image super resolution framework based on compressive sensing and self-learning is proposed in this paper. An over-complete self-learning sparse representation dictionary is trained from the input low resolution image and its down-sample version. Super-resolution under compressive sensing framework does not require the correspondence relationship between the high and low resolution dictionary, which makes proposed method immune to the variety of down-sampling model and show better robustness.In addition, a self-similar block matching method is proposed. This method directly uses self-similar block as the result of first step reconstruction, and then rebuilds the residuals, which preserves the details of the image’s edge well.For video sequences, two compressive sensing super resolution methods are presented:accurate sub-pixel motion estimation method and similarity search method. Accurate sub-pixel motion estimation method combines block matching motion estimation and optical flow motion estimation to get accurate sub-pixel motion vectors. Thus inter-frame similarity information and its accurate position information are computed. Finally, optimal sparse representation equations are got through the multi-frame information, and super-resolution result is obtained by Iteratively Re-weighted Least Squares(IRLS) method. Similarity search method uses similar block matching algorithm to get redundant information of multi-frames. And the mapping relationship of this information and the high-resolution image is computed through bilateral regularization method. High resolution images are constructed through the integration of multi-frames IRLS reconstruction algorithm. A self-learning sparse representation dictionary updating method for the video sequence is also proposed in this paper. The processing frame’s sparse representation dictionary is obtained by removing the unrelated atoms of previous frame’s dictionary, and adding atoms representing new samples. This method is proved efficient in improving the reconstruction performance and reducing computation time.