Temporal Denoising of High Resolution Video

Abstract

<p>This thesis is concerned with the removal or reduction of noises in high resolution<br />video sequences. Many video denoising techniques have been published in the past<br />two decades) with or without motion compensation. They vary in a wide range of<br />complexity) performance, and implementation cost. Also) many existing video denoisers make simplistic assumptions on noise statistics and motion type, and hence<br />their performance depends on the validity of the assumed noise and motion models.<br />To improve the performance and robustness of existing methods) we propose a new<br />joint spatial-temporal video denoising algorithm that combines multihypothesis interframe motion compensation and directional intra-frame filtering. The algorithm takes into account general compound motions, including both global camera motion and<br />individual object motion(s). An affine motion model is used to characterize the global<br />camera movement, whereas a blockwise translational motion model is used to approximate local object motions. Quadtree data structure is used to organize and speed up the computations of block-based motion estimation. Quadtree-structured diamond search is conducted so that a large area can be examined in motion estimation at a low computational cost.<br />In order to achieve the best possible visual quality we augment motion-compensated<br />temporal interframe denoising operation by an intra-frame denoising operation of<br />adaptive directional filtering. The directional filter is designed for the local signal<br />waveform and noise level, and it has the advantage of effectively suppressing noises<br />without blurring edges.<br />The proposed video denoising algorithm is implemented and tested extensively on<br />high-resolution digital cinema contents. The experimental results demonstrate the<br />competitive advantages of the new algorithm in both visual quality and processing<br />throughput.</p>