Digital Image Stabilization by Independent Component Analysis

In this paper, we propose a novel digital-image-stabilization scheme based on independent component analysis (ICA). The method utilizes ICA and information obtained from the image sequence to deconvolve the ego-motion from the unwanted motion of the sequence. We notice that the motion observed in image sequences captured from consumer electronics such as handheld cameras and third-generation mobile phones is mainly caused by two independent motions: the camera motion (ego-motion) and the undesired hand jitter (high-frequency motion). The extensive and successful application of ICA in both the statistical and the signal processing community has helped us to realize that the independence property of these two primary signals facilitates the application of ICA for deconvolution by maximizing their statistical independence. Sets of estimated local motion vectors of the sequence are introduced to the ICA system for separation. Subsequently, we process the unmixed motion vectors to classify the signals into ego-motion and high-frequency motion. Subsequently, when the permutation ambiguity is resolved, the appropriate sign and energy are assigned to the ego-motion vector, resulting in the stabilized image sequence. Experimental results have shown that, apart from the successful deconvolution of the two different motions, the proposed scheme exhibits superior performance compared to other digital-image-stabilization algorithms.