The Study Of Denoising And Super-resolution Algorithms Based On Dictionary Learning

Recently, dictionary learning has an important role in compute vision. It has been widely used in image denoising, image classification, super representation, pattern recognition and machine learning. This paper introduces basic methods of dictionary and spares coding, especially researches the application in image denoising and super representation. Further more, this paper improves these algorithms.Natural images have rich information. There are both smooth regions and texture regions in one picture. To deal with region diversity, this paper proposes a denoise method based on texture intensity. Firstly measure their texture weights using edge detection filters and then add smooth or texture constraint to dictionary learning methods. Pictures’PSNR can improve about1dB.Nonlocally centralized sparse super resolution method(NCSR), add nonlocal mean information to sparse representation. According to steer kernel regression theory, the pixels near center and edge have more influence than others, and have more weights. Multiplying steer kernel weights to patches pixels, use it in PCA dictionary learning and sparse similarity representation. This algorithm intent to affect the validity of dictionary learning and sparse representation accuracy. According to experiments, NCSR-kernel method’s PSNR improved.