A novel framework for fast video inpainting

Abstract

Video inpainting is considered as a complex problem in the current literature. This paper proposes a fast, efficient and automatic method of video inpainting to inpaint moving objects in the video. The presented algorithm employs the spatiotemporal coherency present in the video frames for inpainting while considering the fact that the background either has periodic motion or it remains stationary. The algorithm does not require any manual generation of the mask. Batch frame based inpainting is proposed to maintain motion information in case of background having periodic motion. A new dissimilarity measure; 3D N-SSD is introduced to find similar frames for frame-based video inpainting algorithms. The proposed algorithm is tested for different background and illumination conditions. We have done speed and quality test analysis by inpainting videos of different backgrounds. Quick execution times and high PSNR values for inpainted videos show effectiveness of our algorithm.