| Digital images are inevitably corrupted by noise, which will affecttheir visual qualities and post-processing. A lot of researchers have made agreat contribution to image denoising and proposed some classicalalgorithms and their improvements. But all these methods have somelimitations and cannot get the desired results.Non-local means(NLM) was proposed by A. Buades et al. in2005,which has been proved that can get better results than conventionalalgorithms in image denosing. However, unlike other algorithms, NLMuses a weighted average of all the pixels in the image to recover everypixel, so it is extremely complex. What’s more, many dissimilar pixels areused in weight average will degrade the quality of the denoised image.Some improvements have been proposed to overcome NLM’sshortcomings, among which one of the most important is pre-classification.Inaccurate classification cannot be avoided when using pre-classificationthat has bad influence on denoised results.To reduce the influence of inaccurate classification, this thesisproposes an improved pre-classification method which using the varianceof the whole similarity window and means of four parts of the similaritywindow on the basis of methods of M. Mahmoudi et al. and P. Coupé et al.Accurate pre-classification can eliminate dissimilar pixels from theweighted average which will accelerate the algorithm and improve thedenoised result. At last, we did two experiments to demonstrate theproposed method. The first experiment compares the original NLM, NLMusing pre-classification of P. Coupé et al.(called FNLM) and the proposed method(called ENLM) through peak signal to noise ratio(PSNR) and time.Experimental results indicate that the proposed method can get the highestPSNR in all noise levels and it is faster than original NLM. The secondexperiment compares total variation, original NLM, FNLM and ENLMthrough visual quality and method noise. Its results indicate that methodnoise of total variation and NLM contains many details of original imageand the result of FNLM still have some noise. In contrast, ENLM can wellpreserve details while remove noise. In summary, the proposed method canget better results in less time than original NLM. |
