Image Restoration Based Non-local Constrant And Example Leanrning

As the main way of getting informantion of human is visual system which based on images, the image quality will directly affects human’s understanding of the objective world accurate or not. The presence of the relative motion, atmospheric interference, defocusing, noise and many other factors, led to a decline in images, so image restoration is particularly significant. This technique is aimed to remove blur and noises, while holding the original images without losing details as possible as we can. Image restoration is an ill-posed inverse problem and a very challenging task. This paper studies the models of degradation and restoration, and analyses the common degradation functions, restoration methods and evaluation. Based on above theories, we propose three novel restoration algorithms based on TV (Total Variation) model, NL (Non-Local) means filter, and example learning methods as follows.(1) A novel TV based image restoration algorithm with NL self-similarity constraint is proposed. TV based image deblurring methods are effective in restoring image structures from blurred images. However, they tend to over smooth the local image details. We overcome this drawback by combining local TV regularizer with NL self-similarity constraint, which helps to sharpen the image edges and restore the fine details. The local TV model and NL self-similarity constraint are complementary to each other, making the proposed approach highly effective in removing the blur while preserving image edges.(2) A TV based image restoration algorithm with NL self-similarity constraint with improved weights is proposed. For the problem of the previous method is not ideal for noise suppression, we improve the calculation method of weights in NL means filter by combining that in original NL means filter with Bayesian NL means filter. The improved weights calculation method retains the advantages of both two methods. Experiment results show that the improved algorithm has better performance and effect in image restoration superior to others.(3) An image restoration algorithm based on example learning is proposed. We introduce example learning method into image restoration field using preprocessing technology and propose a novel tri-database model:blurred(B) database, clean(C) database and high-frequency(H) database. We use the technology of patches matching to estimate a low-and middle-frequency image, and a high-frequency image, and then the final restoration image is obtained by adding up the two estimated images. The proposed method can effectively restores the significant details such as textures of objects and the strong edge components in images from training images while reducing undesirable artifacts by averaging overlapped regions between adjacent patches. To reduce the computational complexity, a multi-classification method is also provided for blurred database. The experimental results validate the performance of the proposed approach in both evaluation indexes and visual quality.This research is supported by the National Natural Science Foundation of China (No.61173092), the Fundamental Research Funds for the Central Universities (No. JY10000902045) and the Fund for Foreign Scholars in University Research and Teaching Programs (the111Project)(No. B07048).