Analysis of Optical Flow Algorithms for Denoising

When a video sequence is recorded in low-light conditions, the image often become noisy. Standard methods for noise reduction have difficulties with motion. But the interesting parts in a video is often the ones that are moving, for instance a burglar captured in a surveillance video.

One approach for denoising video sequences is to use temporal filtering controlled by optical flow, which describes how pixels move between two image frames. Today, there exists few studies comparing how different optical flow algorithms perform on noisy video sequences. Four different algorithms have been analyzed in the thesis. Moreover, a comparison on how well they can be used to improve the result of a temporal noise filter has been done. The conclusion of the comparison is that optical flow is useful for noise reduction. Algorithms based on patch matching and edge consistency perform better than algorithms based on color consistency.

A recommendation for future work is to combine the best parts of each algorithm to develop a new optical flow algorithm, specialized on noisy image sequences. Furthermore, develop and implement a sophisticated optical flow based noise filter in camera hardware.