On The Sparse Representation And Compressive Sensing Based Super-resolution Image Processing Technique

Super-resolution (SR) technique provides an effective signal processing method to restore the high resolution (HR) image from either a single low resolution (LR) image or multi low resolution images of the same scene based on some a-priori knowledge. Since SR technology can significantly enhance the quality of the image without need in improving the image sensor hardware device, it is of great importance in not only theoretical analysis, but in practical application. Moreover, SR technology has a wide range of potential application scenarios. The SR technique from single LR image is focused on in this thesis.Some of the existing SR processing algorithms are briefly reviewed. The sparse representation theory which is widely used in image processing field will be introduced. The learning based SR technique which is a hot research topic achieved a better performance than conventional SR technique. The sparse presentation based SR technique is one of the most outstanding techniques. Based on the analysis of the existing joint-dictionary based SR reconstruction algorithms, an adaptive multiple sub-dictionary SR algorithm is discussed in this thesis. In the multiple sub-dictionary based reconstruction algorithm, images are firstly clustered according to their inherent structures, and the suitable sub-dictionary can be adaptively selected from the closest clustering space during the SR image reconstruction procedure. Simulations are presented to validate that better SR reconstruction image can be achieved by using the adaptive multiple sub-dictionary based reconstruction algorithm than by using the conventional joint-dictionary based one.Moreover, the application of three models in SR image processing are discussed, namely, the Adaptively Reweighted Sparsity Regularization model, Iterative Back-Projection model and Non-Local Structure Similarity Model. Simulation results validates that, by employing extra local and global regularization constraints, the SR reconstruction image quality can be further improved.In this thesis, the CS-based SR technique is studied. On one hand, this kind of technique can be employed to reduce the measurement cost and the transmission bandwidth requirements, which is highly desirable for wireless communication, on the other hand, HR image can be restored by using the SR technique. By introducing the Gaussian ambiguity filtering, the un-correlation between the CS measurement matrix and the image basis is increased, the feasibility of the CS based SR technique can be effectively improved. The experimental results are presented to show the viability of the CS based SR technique. The advantages of the proposed CS based SR technique are multi-fold. Firstly, the dictionary learning step can be saved now; Secondly, the CS based SR technique can achieve a reasonable tradeoff in between the time cost and the reconstruction quality.