Gradient-Controlled Gaussian Kernel for Image Inpainting

Image inpainting is the process of filling in damaged or missing regions in an image by using information from known regions or known pixels of the image. One of the most important techniques for inpainting is convolution-based methods, in which a kernel is convolved with the damaged image iteratively. Convolution based algorithms are very quick, but they don’t have good results in structures and textural regions and result in blurring. The kernel size in the convolution-based algorithm is a critical parameter. The large size results in edge blurring, and if the kernel size is small, the information may not be sufficient for reconstruction. In this paper, a novel convolution-based algorithm is proposed that uses known gradient of the pixels to construct a convolution mask. In this algorithm, the kernel size is controlled by the gradient of the image in the known regions. The algorithm computes the weighted sum of the known pixels in a neighborhood around a damaged pixel and replaces the value in the place of that damaged pixel. The proposed algorithm is fast and results in good edges and smooth regions reconstruction. It is an iterative algorithm and its implementation is very simple. Experimental results show the effectiveness of our algorithm.