Study Of Image Registration Based On Affine Invariant Featrues And Optimization Techniques

Image registration is widely used in many application areas such as pattern recognitionand image stitching. The problems and conditions are very different in various types ofapplications. The big change of camera position induces an apparent deformation of theobject image, and the traditional registration algorithms’ performance drops obviously. Thematching of feature points and rejecting bad matches which are related to registrationtechniques are immediately influenced on registration’s performance. This dissertationconcentrates on the image registration algorithm based on the affine invariant features andtechniques related on image registration, in expectation to get well registration undersubstantial viewpoint changes and realize faster and more accurate image registration. Themain work of this dissertation is as follows:(1)The main feature detectors and descriptors are introduced, and their performances areanalyzed systemically. The experiments analyze the repeatability and the number ofcorrespondences of different feature detectors under various transformations. This dissertationevaluates the performance of point descriptors using ROC(Receiver OperationCharacteristics).(2)This thesis proposes Affine-SURF algorithm to get well registration under substantialviewpoint changes. This algorithm simulates the parameters of camera position and gets thesimulated images. Then the feature points of the images are detected by SURF algorithm. Theexperiments prove that Affine-SURF algorithm is fully affine invariant, and it has goodperformance under the viewpoint changes, scale changes and blur.(3)The randomized k-d tree algorithm and hierarchical k-means tree algorithm used inmatching of feature points are introduced. The cost evaluation formula is proposed by thisdissertation based on the performance analysis of linear searching and both previous methods.The time cost and accuracy are very different between these matching methods. Through theperformance analysis and cost evaluation formula, we can choose the best algorithm andparameters for different applications. (4)The Hough-RANSAC algorithm is proposed based on the analysis of the LMedS andRANSAC which are used to reject bad matches. The experiments show that Hough-RANSAChas excellent performance, fast calculation speed and high accuracy when the percent of badmatches is high.