Research On Digital Image Stabilization With Combined Motion

The digital image stabilization is a video jitter removed technique based on imageprocessing. Compared with mechanical and optical stabilization systems, it offers anumber of advantages, including hardware independent, high portability, maintenancefriendly, and capable of stabilizing online video and achieving post processing stepwhich cannot realized by other stabilization methods. These advantages enable thetechnique to be a potential application in many fields, from traffic, consumed device toaerospace, military surveillance and many other fields. With the improvement ofintegrated circuit technique, digital image stabilization is showing excellentperformance and this result promote its widespread application in turn. This dissertationstudies the fundamental principles of digital image stabilization algorithm, analyzes thedefect in main part and gives improved methods accordingly.The dissertation firstly model image motion under camera free movement and thenestablishes similarity equation. Analyzes motion estimation, motion filtering and imagedistortion successively, summarizes the classical algorithms and conclude theircharacteristics.In the third chapter, local motion and global motion estimation are researched.Based on block-matching local vector estimation, a new search method named centerbased extension search algorithm is proposed, the experiment results show its highperformance in various displacements of frames. FAST algorithm is employed to getlocal vectors by tracking feature points, and sets up a self-adaption mechanism to keepfeature numbers in a specified scope. After that, selects reliable local vectors as innerpoints by RANSAC and calculate global motion vectors with LSM.In chapter4, the dissertation studies the motion filtering and image distortionproblems. Aims at the implementation of Kalman Filtering in digital image stabilization,considers velocity state equation and analyzes the results of different Gaussian noiseconstant variance. The Sage-Husa with forgetting factor algorithm is considered torenovate noise covariance matrix, does somewhat approximation in order to avoiddivergence which is caused by the updated negative definite noise covariance. In the image distortion stage, extends adding method from image rotation transform to imagedisplacement, rotation and scale warp. Subsequently, presents image compensationmethods and PSNR evaluation algorithm.