A practical video denoising method based on hierarchical motion estimation

In this paper, we focus on surveillance video denoising with a low-complexity method based on hierarchical motion estimation (HME). The basic idea is to track matching blocks and filter along motion trajectory. The proposed HME method proceeds from large blocks to small blocks and gets motion vectors from coarse to fine. A novel `Y´ shape detector is employed to speed up searching procedure in the top layer where motion vectors from temporal and spatial neighbor are adopted as candidates with different weight. Moreover, new and isolated motions will be caught by random candidate vectors from an offset look up table (LUT). A refinement to deal with discontinuity inside the block is carried out in the intermediate layer and a filter to eliminate singular vectors is used in the bottom layer. Hierarchical process can exclude irrelevant motion vectors and guarantee true motion estimation. It constraints searching range greatly and avoids sophisticated transform comparing with popular denoising methods such as nonlocal means and BM3D. The proposed method can be realized in hardware easily due to limited and reliable searching. Experimental results demonstrate that this low complexity algorithm achieves satisfying denoising performance particularly for surveillance videos under bad illumination.