Multi Stage Directional LMMSE Estimation Based Video Denoising for Poisson Noise: A Frame Localized Frame Based Approach

摘要

In multimedia technology, video processing is a challenging task when it comes to deal with random noise such as poisson noise which is highly pixel dependent. A video affected by such noise shows unpleasant visual effects and thus it is difficult to access the true contents of the frames. The cause of photon noise is due to the discrete nature of photons which adversely affects the signal to noise ratio and amounts to loss of features. Many techniques in the past literature are able to reduce the effect of poisson noise at the cost of edge loss. The work proposed in this paper aims to improve the perceptual quality of each individual frame by reducingnoise and also preserving the contents over the edges as much as possible. A compromise is made between signal to noise ratio and the quality with respect to features in the frames. The advantage of high correlation between adjacent frames is effectively used for estimating the near approximate value of the individual pixel. The current frame (N) blocks (3x3) is localized in the successive frame (N+1) using a novel Block matching technique over a 7x7 window in the successive frame. No thresholding has been considered for block matching. Thus a new localized frame is obtained from the successive frame with reference to the spatial coordinates of the current frame. The Linear minimum mean square error estimate using directional denoising is used over the average of denoised current frame and the denoised localized frame. Directional denoising uses simple or local statistics and thus reduces the computational complexity of the proposed method. Three estimates are evaluated in horizontal/vertical, diagonal and center directions of the noisy pixel and finally fused for final estimate. It is seen that when directional denoising is used over single noisy frame, the signal to noise ratio is better but it is unable to preserve the edges and a blurred image is seen for each frame. Also it is unable to preserve the chroma when applied on individual R, G and B frames. Whereas, when directional denoising over the average of two denoised frames (the current and the localized frame) is used, it is able to improve the video quality by preserving edges and maintaining a low but nearer value of signal to noise ratio to the former case.