Research Of Image Registration Algorithm

Image registration is a fundamental problem in many fields, such as remote sensing, medicine, computer vision and so on. It is used to match two or more pictures taken, for example, from different sensors, or at different times or from different viewpoints of the same sensors.Ⅰ. Image preprocessingThe image color in nature has been analysed, there are so many difference colors that it is difficult to comparation. Transform true color image to gray image has been put forward. Generally there are four methods to transformation. The paper use weighted average method to implement. Experiment results show that images transformed by this method are more close to the person’s subjective feeling on the vision.Ⅱ. Image registration based on grayThe traditional normalized cross-correlation and phase correlation have been analysed, Considering the above two methods are available in translation change but not available in rotation changing image registration, the Radon transformation combined with phase correlation has been put forward. Firstly, it transforms rotate difference to translation difference, then uses phase correlation for registration. Experiment results show that the method is accurate and robust.The common image registration method--the log-polar transform method has been analysed, Considering log-polar coordinates transform exists problem of the uneven sampling:one is oversampling as near the center of the image which causes the computation is large than it is, the other is downsampling as near the edge of the image which causes the miss of usable information. An adaptive polar coordinate transform method to achieve images from a cartesian coordinate system of polar coordinate conversion to sampling uniformity has been put forward. Proved by the experiment, the method is high precision, good real-time in this paper.Ⅲ. Image registration based on feature pointsThe common two feature detection method:Harris algorithm and SIFT algorithm in have been analyzed. Considering the Harris operators extract feature point may focus on a small region, the biggest statistical filtering on image non-maximal suppression method to control the distribution of the angular point has been put forward. Proved by the experiment, the feature points detected by the method are even. For the registration of feature points, firstly, LTS Hausdorff distance has been used to match feature points, remove false matching of feature points with the help of a random sample of consistency. Experiment results show that this method realizes the precise image registration.