A Motion Adaptive Wavelet-Denoising for Frame-Rate Up-Conversion Şükrü Görgülü 1, a , Ömer Nezih Gerek 1, b 1 Anadolu University, Electrical-Electronics Engineering Department, Iki Eylul Kampusu 26555 Eskisehir, Turkey a sgorgulu@anadolu.edu.tr, b ongerek@anadolu.edu.tr Keywords: frame rate up-conversion, wavelet zerotree, symmetric optical flow, denoising Abstract. This study introduces a frame-rate up-conversion method that uses a temporal wavelet zerotree-based shrinkage algorithm over motion trajectory of a video obtained by optical flow. The method starts by optical flow estimation for predicting initial estimates of inserted frame pixels. Then, the predicted frame pixels are denoised using a specific wavelet-based algorithm, where each pixel location is examined independently through its own temporal motion path. The denoising was performed by shrinking zero-tree footprints to remove temporal oddities. The resulting video was observed to have more fluent temporal flow as compared to optical flow - only interpolation. 1 Introduction Evolution of display hardware and bandwidth in recent years brought out very high refresh rates and higher frame rates in videos. Following modern display capabilities, the recent ITU Recommendation BT.2020 announced that new-coming broadcast systems will support higher frame rates up to 120 fps (progressive). However the aged worldwide broadcasting standards support 24-30 fps and there are a huge amount of old-standard video recordings. Playing these videos in new systems needs suitable frame up-conversion (FRUC) techniques while keeping the playback quality at acceptable merits in terms of motion blur. In this paper, a wavelet-shrinkage based denoising algorithm is proposed to be embedded motion-adaptively into a FRUC method. The study combines wavelet zero-tree analysis with optical flow estimation (OFE) based frame interpolation. The process starts with temporary interpolation of new inter-frames using 2D motion vector (MV) information gathered from OFE. The new frame sequence including interpolated frames, and MV information are employed together to extract each pixel's motion route. The proposed wavelet shrinkage algorithm runs over these extracted motion paths. The paper explains detailed parts of the processing pipe shown in Fig.1. The literature contains several studies on the FRUC topic. Following studies utilize wavelets for efficient interpolation of frames. Guo et al. propose a denoising scheme in temporal direction based on multi-hypothesis motion compensation [1]. Their method was derived by commonly used three dimensional (3D) filters to operate only in one dimension, which makes the process much faster than its ancestors with less memory requirements while giving better results. Zlokolica et al. studied on a technique that first performs recursive temporal denoising through the predicted motion path and then applies appropriate spatial filter as reported in their paper [2]. Rahman et al. proposed a joint probability density function for modeling the video wavelet coefficients of any two neighboring frames in their study [3]. Using that statistical model, denoising was performed. Cheong et al. proposed an adaptive spatio-temporal filtering, which combines wavelet decomposition, Wiener filtering, and block based motion estimation and compensation techniques in their study [4]. Rajpoot et al. present a study that utilizes an optimal 3D wavelet transform of video and performs a modified form of the Bayes-Shrink thresholding method to suppress the noise for improving restoration of the video [5].