A Multi-scale Evaluation Method for Motion Filtering in Digital Image Stabilization

In digital image stabilization, the real camera motion usually illustrates different modes and properties, such as large moving trends, wandering patterns, diverse jitters and so on, which make the performance of motion filtering very difficult to be universally evaluated. This paper proposes a universal and objective evaluation method for motion filtering in time scale space. According to the wavelet variance, the fluctuating energy of real camera motion is decomposed into multiple scale levels, and then, the energy variations in different scales between the original motion and the smoothed intentional motion are fused, at the same time the energy ratios of high frequency components and low frequency components in the original motion are assessed, which is considered as fusion weights to represent the concern between the path fitness and path smoothness, so the evaluation result can achieve the tradeoff between the filtering consistency and filtering stabilization adaptively according to the specific property of real camera motion. In the experiments, the evaluation results are illustrated to have precise distinguishing ability and consistent with the intuitional observation.