A block-based adaptive three stages motion estimation using early termination for anti-shaking Sangwoo Ahn, Yeonchul Jeung, Kyeongyuk Min, Jongpark Kim and Jongwha Chong Department of Nanoscale Semiconductor Engineering Hanyang University Seoul, Korea ahnsangwoo@hanyang.ac.kr Abstract— In this paper, a block-based adaptive three stages motion estimation using early termination which is used for anti-shaking is proposed. Low computational complexity and accurate motion estimation are demanded to anti-shaking for mobile phone. Therefore three stages which consist of the predicted motion vector (PMV), the two-bit transform using second derivatives (2BT-SD) and the two steps search (2SS) are performed stage-by-stage. The three stages are early terminated up to result of stage. So the trade-off between computational complexity and accuracy is mediated by the characteristics of early termination which is determined by characteristics of frames. To evaluate the proposed algorithm, peak signal-to-noise ratio of reconstructed frames and current frames is measured. And the proportions of each stage are used to evaluate computational complexity. Keywords-component; fast motion estimation; block motion estimation; predicted motion vector; two steps search; two-bit transform; anti-shake; I. INTRODUCTION In mobile phone camera, anti-shaking technology is needed. Conventional digital camera uses optical method of anti-shaking. But hardware cost and space are limited in mobile phone. Therefore software approach is applied to mobile phone. It is important technique for the motion estimation to anti-shaking by S/W [1]. But motion estimation requires a high computational cost of the total computational cost of anti-shaking. Therefore, it is important to reduce computational cost of motion estimation. To reduce complexity of motion estimation, various kinds of techniques are proposed. First category is related with search range or search patterns. These kinds of method search blocks that have high probability of matching instead of all blocks of the image. So, these kinds of method reduce the amount of computations and the time of computations. Three-step search (3SS) [2], new three-step search, cross- search, polynomial search (PS) [3], diamond-search (DS) [4] and four-step search (4SS) [5] are examples of the methods falling into this category. Second category is related with comparing method. In motion estimation, it is very important to determine which block would be well matched with reference block. So, the methods that measure the distortion level between reference block and current block are needed. Generally, mean of absolute difference (MAD) is used. But it has very high computational complexity. So techniques of second proposed approach are simple but high accurate methods. One-bit transform (1BT) [6], two-bit transform (2BT) [7], two-bit transform using second derivate (2BT-SD) [8] are popular examples of methods falling into this category. In 1BT, points of images are set to 0 or 1 according to mean of block. And number of non-matching point (NNMP) is computed between blocks of neighboring frames. NNMP is used as the distortion factor. So, block which has minimum NNMP value is selected as matched block. The 1BT is simple, so computational complexity is low but makes inaccurate results. And the 2BT is proposed to improve accurate. The 2BT uses two bit planes. First bit plane is same as the 1BT. And second bit plane is set to 0 or 1 according as mean and standard deviation. In the 2BT, NNMPs are computed using first bit plane and second bit plane, so the 2BT can separate the set of compared blocks more accurately. And the 2BT- SD is proposed for more improved accuracy. In the 2BT-SD, second bit plane is discrete to 0 or 1 according as mean, positive second derivative and negative second derivative that contain characteristics of the image. To reduce computational cost and maintain accuracy, the trade-off between computational complexity and accuracy must be considered. Therefore, the 2BT-SD of second category which reduces computational cost much, 2SS of first category which has high computational cost but accurate method and PMV which reduce redundant computation are applied to the proposed algorithm. These three algorithms are applied stage-by-stage and early terminated by result of stage. Therefore the proposed algorithm mediates trade-off according as the characteristics of frames. The proposed algorithm performs motion estimation based on the PMV. And values of distortion are measured between reference block and block estimated by the PMV. If the value of distortion is larger than the threshold, the two-bit transform using second derivative is processed. And then values of distortion level are measured between reference block and block estimated by 2BT-SD. If the value of distortion is larger than another threshold again, the two steps search is processed [10]. This paper is organized as follows. The Section explains the proposed algorithm and the Section explains result of the proposed algorithm. 2011 Fourth International Joint Conference on Computational Sciences and Optimization 978-0-7695-4335-2/11 $26.00 © 2011 IEEE DOI 10.1109/CSO.2011.313 336