Image Denoising Research Based On Kernel Regression And Non-local Method

The kernel regression analysis is the latest development of the traditional regression analysis, it is widely used in the domain of image reconstruction, super-resolution, data mining and so on. Non-Local method is the popular denoising method at present, it considers the global features, and its result is better than other methods, however, it is quite slow. This thesis is based on the national natural science fund project, we combine the upper two methods and use it to denoise the image.We firstly introduces the concept of image processing and the common noising model, then we review the classic image denoising methods, we find that the classic denoising methods have a disadvantage that they rely on a specific model of the signal, it restricts the scope of these methods because only some of the real conditions are fit for the specific model and you can never change the model. The classic kernel regression method improves this weak point, however, it ignores the local characteristics, we improve the classic kernel regression by considering the local information.The kernel regression is a local method which ignores the global pixel relations. This thesis introduces the non-local method, it use the global features to calculate the weights, which generates excellent results when the image has many similar features. Then we establish its relation with the kernel regression methods, and we find that the local model of the non-local method is a 0th order kernel regression model, thus we extend the non-local model to high order and improve the results.Finally, compared to the traditional methods, we have got better results which reserve the local texture and structure information of the image through experiments and data analysis, that's because the combined method consider both the global and local features.