An image inpainting technique based on nonparametric kernel estimation

Image inpainting is an important research topic in image processing and computer graphics, etc. Its objective is to restore the lost information or remove the unnecessary object according to around image information, which can be used to restore old photo, remove text and conceal errors in videos. However, previous algorithms are not to be perfectly performed. In this paper, a new image inpainting algorithm based on nonparametric kernel estimation is presented. From a statistical point of view, the inpainting can be viewed as an estimation problem with the missing data. Firstly, an initial value was estimated by using the known information as Taylor expansion at the missing point. And then a cost function was obtained by utilizing a Gaussian kernel function to penalize the estimate error. Finally, the missing data are restored through minimizing the cost function. Experimental results show that the proposed method can mend a large variety of images effectively and robustly.