Abstract

Video completion refers to automatically restoring damaged or removed objects in a video sequence, with applications ranging from sophisticated video removal of undesired static or dynamic objects to correction of missing or corrupted video frames in old movies and synthesis of new video frames to add, modify, or generate a new visual story. The video completion problem can be solved using texture synthesis and/or data interpolation to fill-in the holes of the sequence inward. This thesis makes a distinction between still image completion and video completion. The latter requires visually pleasing consistency by taking into account the temporal information. Based on their applied concepts, video completion techniques are categorized as inpainting and texture synthesis. We present a bandlet transform-based technique for each of these categories of video completion techniques. The proposed inpainting-based technique is a 3D volume regularization scheme that takes advantage of bandlet bases for exploiting the anisotropic regularities to reconstruct a damaged video. The proposed exemplar-based approach, on the other hand, performs video completion using a precise patch fusion in the bandlet domain instead of patch replacement. The video completion task is extended to two important applications in video restoration. First, we develop an automatic video text detection and removal that benefits from the proposed inpainting scheme and a novel video text detector. Second, we propose a novel video super-resolution technique that employs the inpainting algorithm spatially in conjunction with an effective structure tensor, generated using bandlet geometry. The experimental results show a good performance of the proposed video inpainting method and demonstrate the effectiveness of bandlets in video completion tasks. The proposed video text detector and the video super resolution scheme also show a high performance in comparison with existing methods.