Research On Image Registration Based On Feature Points

Image registration is a key technique in pattern recognition and image processing,and it is widely used in many application areas such as computer vision, remotesensing, image fusion, image super-resolution reconstruction and medical imageprocessing. The requirement for the accuracy, adaptability and efficiency of imageregistration method increases along with the development of these applications. Thereare two main categories of image registration according to the method: area-based andfeature-based. The feature-based image registration methods translate the analysis ofthe whole image to the one of some feature, so they are very efficient. Most of thesemethods are robust to change of gray value, image distortion even occlusion. Thisdissertation concentrates on the image registration methods based on the featurepoints in expectation to realize the fast and accurate image registration under compleximaging conditions. The research involves four aspects:(1) The main related techniques of image registration based on feature points areintroduced and analyzed systemically. These consist of the extraction, description andmatching of feature points, the estimation of the spatial transform matrix, imageresampling and interpolation and the analysis of the accuracy of image registration.And the former three parts is more important, so it enumerates some classicalgorithms with regard to them and the strongpoints and shortcomings about thesealgorithms are analyzed. (2) Starting with the concept of human visual neuron's receptive field (RF), ahyper-complex neuron and its mathematical model (scale-interaction model) ispresented. The Gabor wavelets are replaced by the Mexican-hat wavelets in the modelto generate the scale-interaction of Mexican-hat wavelets. The scale-interaction ofMexican-hat wavelets could extract close number of feature points that correspondingto the same location in the images distorted by rotation,brightness, blur and noises.For the problem that the feature points extracted are corresponding to the differentlocations in images of different scales, the method that resizes the Mexican-hatwavelets by the scale factors is presented. Then,Log-Polar Transform and normalizedPseudo-Zernike moments are relatively used to match these feature points and tocomplete the registration for different image types. The experiment indicates that theproposed algorithms extract feature points and match them exactly and eliminatewrong matched points effectively and achieves pixel precision registration result.(3) A fast and robust image registration method based on the SURF-DAISYalgorithm and randomize kd trees algorithm is presented. Firstly, a SURF-DAISYalgorithm is presented. In this algorithm, feature points are extracted using classicSURF (Speeded Up Robust Features, SURF) feature detector in the reference imageand sensed image respectively. Then, DAISY descriptors are utilized to substitute forthe customary SURF description algorithm to characterize those feature points tospeed up the process. The strategy exerts the robustness of SURF feature detector andthe efficiency of DAISY descriptor adequately. After that, the matching process iscarried out via a multiple randomized kd trees algorithm and the RANSAC (RandomSample Concensus,RANSAC) is used to eliminate wrong matches. At last, the besttransform parameters are estimated by the least square method and the imageregistration process is accomplished. The experiments indicate that the proposedmethod reduces the time cost of the registration process up to45.6%compared to theclassic SURF algorithm while nearly retaining the performance, and it proves that therecommended algorithm is fast and robust.