Gaussian mixture model based phase prior learning for video motion estimation

Optic measurement nowadays shows more prominent features over the traditional sensor-based measurement techniques as it does not need sensor installment and can provide simultaneous measurements with very high spatial resolution. Phase-based video processing has been widely used in optic measurement for capturing small motions. It relies on the assumptions that the variance of noise has to be much lower than the pure signal and small motions are encoded in the phase shift of an individual pixel's color which fundamentally restricts its usefulness. To tackle these issues, this paper proposes a novel perspective of phase estimation by utilizing the image priors on phase patches. The patch group based Gaussian Mixture Model (PG-GMM) learning algorithm is used to learn the nonlocal self-similarity (NSS) prior from training images. Then the phase information is modeled by learned patch group priors, and further optimized using the steps of gaussian component selection and weighted sparse coding. The proposed method is validated in both magnification and modal analysis applications in the example analysis. Compared to the phase-based method, the proposed method achieves high-quality magnifications on the real video signal, with fewer artifacts and better anti-noise performance. Clearer time domain motion estimates of video components can also be extracted by the proposed method.