The Research On Image Super-Resolution Reconstruction Algorithm Based On Sparse Representation

The 21 st century is the era of information technology, the wide application of multimedia technology and the widespread popularity of digital devices is one of the characteristics of this era, digital image acquisition, storage and forwarding is an integral part of this feature. However, with practical application, due to the restrictions of image acquisition devices, other hardware devices, environmental interference and motion blur and other conditions, the image resolution can not reach the expected effect. Therefore, it is necessary to develop the image reconstruction algorithm to enhance the resolution of the low resolution image, which will be of great significance to the development of multimedia technology.Among the many reconstruction algorithm, sparse representation of super-resolution reconstruction algorithm has attracted much attention because of its unique advantages. The algorithm is based on sparse representation theory, with single input low resolution image, and any image block of image can use a pre-trained low resolution dictionary to obtain the sparse representation coefficients, then using the consistency principle of sparse representation coefficients obtained from the high and low resolution dictionary and the high resolution dictionary to get high resolution image reconstruction. In this dissertation, the research base on this study, the algorithm improve the existing problems.To begin with, proposed image feature extraction process improvement strategies: improved first and second order filter operator. The applicability of the sparse representation coefficients calculated by the image after feature extraction is enhanced, it can effectively remove useless information. In this dissertation, the high frequency detail of reconstruction image is more delicate by the improved feature extraction operator. Matlab simulation results show that the PSNR value of the restored image by the improved feature extraction operator is improved 0.1025 d B compared with the original sparse representation algorithm.Next, an improved algorithm for the initial estimation of the reconstruction algorithm is proposed. Considering the directional edge of the image,use the edge direction interpolation as the initial interpolation. Edge direction interpolation makes full use of high and low resolution images of geometric duality theorem. By comparing PSNR value, compared with the original algorithm increases 0.2825 dB.Once again, the improvement strategy of the search window in the reconstruction process is proposed. The original algorithm does not deal with the pixels around the image with sparse restoration algorithm, but use the bicubic interpolation. In this dissertation, with 4×6、3×8、2×12、1×24 image block trained four dictionary sequentially traverse in turn again to ensure that all pixels were sparse recovery pixel traversal balance. The experimental results show that compared with the original algorithm the PSNR value is improved by 0.1830 dB.Besides, improvement of iterative end sign of iterative back projection algorithm. In order to effectively control its convergence, this dissertation uses the relative error as a sign of the iterative results. Compared with the original algorithm, the improved algorithm can get better convergence results.Finally, in this dissertation, the improved algorithm is applied to the restoration of the blurred license plate, reconstruction results show that the number of fuzzy license plate can be recovered.