New digital cameras, such as Canon SD1100 and Nikon COOLPIX S8100, have an Auto Exposure (AE) function that is based on motion estimation. The motion estimation helps to set short exposure and high ISO for frames with fast motion, thereby minimizing most motion blur in recorded videos. This AE function largely turns video enhancement into a denoising problem. This paper studies the problem of how to achieve high-quality video denoising in the context of motion-based exposure control. Unlike previous denoising works which either avoid using motion estimation, such as BM3D [7], or assume reliable motion estimation as input, such as [13], our method evaluates the reliability of flow at each pixel and uses that reliability as a weight to integrate spatial denoising and temporal denoising. This weighted combination scheme makes our method robust to optical flow failure over regions with repetitive texture or uniform color and combines the advantages of both spatial and temporal denoising. Our method also exploits high quality frames in a sequence to effectively enhance noisier frames. In experiments using both synthetic and real videos, our method outperforms the state of the art [7, 13].

BACK