| In process of image acquisition and target recognition, Super-Resolution(SR) reconstruction, a kind of software technology which is beyond current hardware level, has a great market demand in practical application. In the past 20 years, it has become an attractive research hotspot in the field of image processing. Super-resolution reconstruction can be applied to some practical problems in different areas, such as satellite and medical image processing, facial image analysis, text image analysis and so on. So far, many super-resolution algorithms based on different angles and different theories have been presented by scientific research workers.This paper mainly presents a super-resolution reconstruction algorithm for multi-frame images in spatial domain. Firstly, this paper briefly introduces basic theoretical knowledge and development background of SR, and then analyzes several existing robust SR resolution technology for multi-frame images. On this basis, this article mainly researches as following:(1) NLM-SR algorithm based on Bayesian theory. By analyzing experiment results of simulation of original NLM-SR, it is observed that the error in computing the weights of similarity of image blocks makes better super-resolution implemented only after several iterations, and finally leads to low convergence speed. To solve the problem, this paper utilizes Bayesian theory to reconstruct an equation of computing the weights by fitting error curves, which improves the convergence speed. And by comparing with original NLM-SR, algorithm performance has also been improved.(2) NLM-SR algorithm based on fuzzy motion estimation. NLM-SR, with no parameter estimation, greatly deduces the possibility of error in model estimation, but has high computation complexity. To solve the problem, we involve in a concept of fuzzy motion estimation referring to two-step method in sequence image denoising algorithm. Then NLM-SR is able to greatly reduce the search range in block matching. By comparing the experiment results, involving the concept does not impact the SR performance of NLS-SR, however, it has greatly improved algorithm efficiency.(3) Bayes NLM-SR algorithm based on fuzzy motion estimation. Combining above two improved algorithms, this paper presents a novel Bayes NLM-SR algorithm based on fuzzy motion estimation, called ME&BNL-SR. Compared to original NLM-SR, this algorithm has great improvement both in visual effect and algorithm performance. Besides, in order to apply presented algorithm to practice, super-resolution reconstruction of sequence images of license plate has been implemented with experiments. Experiment results on composite and real images both show that ME&BNL-SR is effective on super-resolution of license plate.At last, this paper summaries encountered problems in the process of researching super-resolution and gives an outlook of future super-resolution reconstruction algorithm. |
