Kernel Regression Image Denoising Based On Non-parametric Estimate

In the course of processing image data, the quality of the image may be damaged because of many reasons, such as noise. As to the further management, it's necessary to de-noise the image. In recent years, the kernel regression method, which is from the non-parametric estimate theory, gains great development, and has been penetrating into each field and achieved some success in image de-noising. Although there are many image de-noising methods, using kernel regression to de-noise image is still a remarkable problem, and it's very significant to research in theory and application.In this thesis, comprehensive research has been done for the image de-noising based on the kernel regression. The research contents are follows:Firstly, image de-noising knowledge and some popular de-noising technologies are summarized.Secondly, the regression theory and three primary regression models are introduced in detail.Thirdly, traditional kernel regression de-noising method is researched and its develop potential is proved according to the experiment result, and then a data adaptive kernel regression algorithm is proposed via combining classical algorithm and characteristic of image. Fourthly, an iterative procedure is introduced into the adaptive kernel regression algorithm, which makes the estimated images have superior quality. Several important properties are illuminated through experiments: 1.With same global smoothing parameter, after the best estimator is reached, RMSE grows as iteration grows; 2.The Bias of the estimator grows as iteration grows, and the contrary with Var; 3.With different smoothing parameters, iterations needed to reach the best estimator are different. The smaller the smoothing parameter, the more iterations are necessary to reach the minimum RMSE; 4.The RMSE of the best estimator with different smoothing parameters are nearly the same, but visual effect is different. Smoothing parameter with more iteration can keep the image details better, give good visual result. Fifthly, a new kernel regression de-noising algorithm combined with sobel operator is proposed, whose feasibility is analyzed in theory, which can highly decrease the computational cost and have better de-noising effect especially at high regression order.Sixthly, how to implement the algorithm in Visual C++6.0 is provided, the comparison between all algorithm in this thesis is provided via experiments, and result shows good efficiency of the iterative algorithm. Lastly the algorithm is applied to medicinal image de-noising, and comparison between popular de-noising methods is provided, and result shows superiority of our algorithm in medicinal image de-noising.Now the achievement of kernel regression de-noising methods not only widens the application field using the kernel regression de-nosing, but also forces the development of these fields. At the same time, all new problems, which are feed back during the application, will enrich the kernel regression content and force the development of the application in image process.