Probabilistic Global Motion Estimation Based on Laplacian Two-Bit Plane Matching for Fast Digital Image Stabilization

Abstract

Digital image stabilization (DIS) is a technique to prevent images captured by a handheld camera from temporal fluctuation. This paper proposes a new DIS algorithm that reduces the computation time while preserving the accuracy of the algorithm. To reduce the computation time, an image is transformed by a Laplacian operation and then converted into two one-bit spaces, called L+ and L− spaces. The computation time is reduced because only two-bits-per-pixel are used while the accuracy is maintained because the Laplacian operation preserves the edge information which can be efficiently used for the estimation of camera motion. Either two or four subimages in the corners of an image frame are selected according to the type of the image and five local motion vectors with their probabilities to be a global motion vector are derived for each subimage. The global motion vector is derived from these local motion vectors based on their probabilities. Experimental results show that the proposed algorithm achieves a similar or better accuracy than a conventional DIS algorithm using a local motion estimation based on a full-search scheme and MSE criterion while the complexity of the proposed algorithm is much less than the conventional algorithm.