Segmentation-Based Video Upscaling without Motion Estimation

Abstract


 This paper shows an effective method for video upscaling or super resolution (SR) without using an explicit motion estimation step. Exploiting the Non-Local Means (NLM) algorithm in order to bypass motion estimation, which is often complicated, our method proposes some modifications to ensure a good compromise between noise cancelling and detail preservation. A detailed consideration of the NLM algorithm is carried out to propose an efficient distance computation and the best eighbors for the reconstruction of each SR pixel. Moreover, efficient segmentation algorithms are also considered to build a novel upscaling framework that is adapted to spatial contrast. The satisfying results with real videos illustrated the advantages of upscaling without motion estimation compared to motion estimation-based upscaling, as well as the role of segmentation in video super resolution.