Self-adapted optimization-based video magnification for revealing subtle changes
Enjian Cai, Dongsheng Li, Hong‐Nan Li, Zhilin Xue
Abstract
Video magnification techniques can reveal subtle temporal variations that are difficult or even impossible to see with the naked eye and display them in an indicative manner. State-of-the-art approaches rely on hand-designed filters or precise prior knowledge of the target signal and produce magnif ied outputs that are always bounded by the spatial support of the related pyramids. This paper proposes a method for adaptively magnifying subtle video changes by directly solving three key optimization problems. To efficiently model the magnification transformation, alternating direction method of multipliers is employed to solve convex variation-detection optimization problems. Following the transformation step, the perturbation problem is innovatively solved using a forward additive iterative approach to iteratively minimize the dissimilarity between the original and magnified sequences based on the enhanced correlation coefficient. The proposed method can be applied to videos overlaid with different types of temporal change to obtain motion and color magnification. Quantitative and qualitative experimental comparison of the proposed method with state-of-the-art techniques reveals that it produces magnified videos with improved visual quality and substantially fewer artifacts or blurring.