Video Inpainting Using Advanced Homography-based Registration Method

摘要

Abstract A new video inpainting technique for videos taken from free moving cameras is suggested in this research paper. The effective results of video inpainting can be achieved by maintaining spatiotemporal coherence while filling the holes in the target frames. This is possible only with the proper registration of source frames to the target frame. Image registration plays a vital role in the process of video inpainting to obtain effective results. An advanced homography-based image registration method is introduced, based on HALF-SIFT: high accurate localization feature for SIFT to extract feature points without localization error. The covariance matrix has been used to estimate the localization error. Further, new inlier selection method using CW MLESAC and refining is carried out for homography matrix with CW L-M. This iteration process can improve the accuracy of image registration. After registering frames to the target frame, the hole is inpainted by globally minimizing the energy cost function. The proposed video inpainting is applied to several complex video sequences. Experimental results are outperformed in visual quality when compared with the state-of-the-art methods. The performance metrics like peak signal-to-noise ratio and Structural Similarity Index are determined and compared with existing methods for different video sequences.