Registration Of Images With Affine Distortion Based On MSER And Phase Congruency

Image registration is an important and fundamental step in many computer fields, such as image fusion, image stitching, object detection and 3D reconstruction. Generally, images shot at different times, from different viewpoints or by different sensors often contain some geometric transformations, such as translation, rotation, scaling or affine transformation. The goal of image registration is to correct the coordinates and the deformations by searching the correspondences between images with the above distortions. The structure change caused by affine geometric distortion is often larger than that caused by translation, rotation and scaling transformation, which makes it difficult for the traditional feature detection methods to search the correspondences between images. Therefore, it is harder to align the images with affine transformation.Firstly, the four fundamental steps of the featured-based image registration are discussed in detail in the dissertation, i.e., feature detection, feature matching, transformation model estimation, and image re-sampling and transformation. Based on this, some commonly-used algorithms, such as Harris corner detection, Scale Invariant Feature Transform(SIFT) based point detection and matching, Affine-SIFT(ASIFT) based point detection and matching, phase congruency based point detection and Maximally Stable Extremal Region(MSER) detection, are studied thoroughly, and some problems of them are analyzed and summarized.Secondly, a novel method based on MSER and phase congruency is proposed to address the registration of images with affine transformation. The proposed method contains three steps:(1) The MSER detection and matching method is performed on the reference image and the image to be registered, respectively. By fitting and normalizing the centroids of the matched MSERs, two circular regions that contain roughly the same image content are obtained. And the coarse affine transformation matrix between the two input images is estimated by the matched MSER pairs.(2) A novel feature point detection algorithm based on the Gabor filter decomposition and phase congruency is presented. Two feature point sets are thus achieved by performing the proposed feature detection algorithm on the two coarsely aligned circular regions, respectively.(3) The affine transformation matrix between the two feature point sets is obtained by using a probabilistic point set registration algorithm, and the final affine transformation matrix between the reference image and the image to be registered is calculated according to the coarse affine transformation matrix and the affine transformation matrix between the two feature point sets.Finally, the proposed method is implemented with Matlab programming in the Windows XP environment. And some other classical area-based and feature-based registration methods on images with affine transformation are also performed for comparisons. Several sets of experiments demonstrate that the proposed method obtains higher registration accuracy than the MI-based and CC-based methods. The proposed image registration method performs better than the SIFT-based image registration method when aligning images with large affine distortions. Compared with the MSER-based image registration method, the proposed method obtains more correct matches, higher feature repeatability rate and registration accuracy. And the proposed method has higher computational efficiency than the ASIFT-based image registration method. In addition, for images with illumination change, blur and JPEG compression, the proposed method also achieves satisfactory registration results.