Video noise reduction using H.264 multi-frame trajectory

Abstract

In video capture, noises such as Gaussian and impulse noise may exist because of insufficient luminance or defect of image sensors. Noise in video will seriously affect human visual perception and reduce compression efficiency. Therefore, noise reduction is a necessary part of video processing. Spatial filter and temporal filter were proposed to reduce noise in video in previous works. Both of them have disadvantages such as ghost effect of spatial filtering and blurring in temporal filtering. Consequently, we propose a spatio-temporal filtering scheme that utilizes motion compensation with multi-reference frames in H.264 to remove noise. It utilizes the variation of inter mode distributions to detect noise and determine the parameters for spatial filter. In the time domain, a reference pixel is selected from multi-reference frames according to the motion trajectory and MSE criterion. The proposed adaptive spatial filter and non-linear temporal filter can effectively remove Gaussian and impulse noises to improve video quality up to 0.83dB and 8.17dB in PSNR respectively. Meanwhile, it can also boost the compression efficiency by reducing the bitrate up to 84.85% in our experiments.