Improved Harris corner detector algorithm for image co-registration M safy,Guang ming Shi,Zhenfeng Li Xidian University Xian, China Msafy_2010@yahoo.com Ahmed saleh amein Military Technical College Cairo, Egypt Abstract— Image registration is the process of removing the offset between two or more images of the same location from different angles. One of the two images is the master image and the other is the slave image. In this paper, an effective algorithm is proposed to align the two images using Harris corner detector. This algorithm is considered as a connection between the area based methods and the feature based methods .It has the advantages of avoiding the high computational cost of the fine co- registration in the based area methods and at the same time it improves the Harris corner detector efficiency. In this algorithm, the cross-correlation function is used to detect the coarse offset between the master image and the slave image, while the fine offset between the two images is detected by the Harris corner detector and the Scale-Invariant Feature Transform (SIFT). Random Sample Consensus (RANSAC) algorithm is used to estimate the transformation model. The proposed algorithm is evaluated by using the coherence map. The results show that the proposed algorithm is more effective than using Harris corner detector alone. Keywords— Cross correlation; Harris corner detector; SIFT; RANSAC; coherence I. INTRODUCTION Image registration is one of the important steps in the development of digital elevation model (DEM). The slave image must be transformed using image registration techniques to match the master images. Because of the increase of the image resolution, image registration needs heavy computational effort. Thus, for image registration high-performance computing techniques are critically needed. There are two main approaches for image registration. One of them belongs to the area based methods and the other to the feature- based methods. In Brown’s survey paper [1] the area based methods applies a cross- correlation measure between the images which needs heavy computational effort (especially in fine co-registration step). Without essential preprocessing the process will be affected by noise. In other papers by Ton and Jain [2] and Alt and Guibas [3] Feature-based methods give more accurate results because features are usually more reliable than intensity or radiometric values. Features may be control points, corners, line segments, etc. There are some disadvantages of Feature-based methods like the following:1) It will not be effective if the feature extraction process yields significant numbers of missing or spurious features 2) It is computationally expensive, especially when large point sets are involved and when the transformation space has many degrees of freedom. In this paper, our work is based on the feature-based methods. But we are interested in the following: 1-Do all the feature-based methods give the same result for the same area? 2-What is the relation between the feature-based methods and the offset between the images? So two feature-based methods are used one of them is the Harris corner detector [4] and the other is the Scale invariant feature transformation (SIFT) (this algorithm was proposed in 1999 by Low and then enhanced in 2002 by Brown and in 2004 by Lowe). This paper is structured as follows: we will start with a short literature review in Section II, which describes our registration procedure in details. In Section III, we give the experimental results and assess the registration performance. The conclusion along with a discussion is given in Section IV. II. METHODOLOGY In this work, the SAR co-registration procedure is divided into three main steps: 1) Find the coarse offset between the master and slave images and mapping the slave image according to this offset 2) Find the sub-pixel offset using one of the feature-based methods 3) Selecting correspondences between Points and removing Outliers using Random Sample Consensus (RANSAC) Algorithm. The co- registration performance is evaluated by coherence factor. Fig. 1. Procedure of the proposed method. The removal of the offset in pixel level between the two SAR images is called the coarse co-registration step. These two images must be taken for the same location and under the same flight conditions. The SAR image located closer to the target of interest is called the master image, while the other is called the slave