Abstract:
Conventional wavelet shrinkage schemes for video denoising lack correlation exploitation amongst neighboring pixel values within a video frame or image. These techniques are applied directly to the images or video frames in wavelet domain without any preprocessing thus do not have any chance to make use of the resemblances that exist among the neighboring pixels.
Applying a suitable Path permutation on functional values of pixel data brings functionally close values closer to each other that facilitate exploitation of correlation amongst neighboring pixels for wavelet shrinkage.
We have developed a method that operates on each pixel level of data array prior to application of transform domain filtering on image sequences. The three dimensions, two spatial and one temporal of a video sequence is processed in sequential manner. The data set is formed by arranging a sequence of noisy videos frame-wise into a one-dimensional data array to exploit the geometry of wavelet sub-bands. The algorithm makes use of spatial coherence within the same frame as well as the temporal coherence between consecutive frames of a video observed at different moments in time. In the first part, a one dimensional wavelet shrinkage is applied along suitably chosen path vectors, while in the second part, temporal averaging intervals are created by pixel based motion detector with recursive, selective averaging. The proposed hybrid algorithm is tested on many benchmark videos with experimental results exhibit that the proposed technique produces better qualitative and quantitative results when compared with conventional wavelet shrinkage methods. Our video denoising technique operates on image sequences degraded by zero-mean additive white Gaussian noise (AWGN).
