An Image Registration Algorithm Based On Hausdorff Distance And Genetic Algorithm

Image registration is one of the basic digital image processing methods, which mainly registers two or more digital images, mostly geometrically, obtained at different time, by different sensors, on different angle of view or on different filming conditions. In recent years, image registration is studied in many different applying domains, so that it plays a very important role in computer vision, pattern recognition, medical image analysis and remote sensing. Image registration has become an essential part during many studies and researches.A key problem in image registration is to find an effective way in which we can evaluate the similarities of the digital images. In 1991, a method based on Hausdorff distance was proposed, which is about the computing of the image similarities. From then on the Hausdorff distance has been utilizing in many experiments and researches, and as a standard of measuring the positions of images, it has been proved to be a good method in registration accuracy. However the naive Hausdorff distance has a defect during image registration, which is the sensitivity of noise and outliers. While as the improved Hausdorff distance, the LTS (Least Trimmed Square) Hausdorff distance can overcome these problems easily. So the accuracy and stability of image registration which is based on LTS Hausdorff distance is more preferable.In this paper, an image registration algorithm based on LTS Hausdorff distance and genetic algorithm is proposed. Before the algorithm begins, the reference image and the unregistered image are processed using image sharpening, image smoothing, image binarization and edge detection. Then on the result of the pre-processing, the processed images are registered using genetic operations, during which the LTS Hausdorff distance is used as the fitness function; finally the goal is to find the best transformation parameters by genetic algorithm. The main contents are as follows:(1)It introduces the LTS Hausdroff distance after explaining the definition, theorem, as well as the problems which may come out during the image registration. Moreover, as the main standard measuring the similarities during the image registration, it can perfectly solve the problems caused by the sensitivity of noise and outliers, which mainly appear in na?ve Hausdorff distance. More importantly it can guarantee the accuracy of image registration.(2)It narrates the whole procedure of the image registration based on genetic algorithms and LTS Hausdorff distance. During the process of pre-solving of the images, the images are processed by the operation of sharpening, smoothing, binarization and edge detection. First, it is because of the methods of Butterworth filter which used in edge detection to get the point features of images that the sketches of the images can show more clearly and accurately. Then the GA is applied to the image registration, and the GA used in the algorithm is improved on the ground of the features of image registration. The improvement including: using real number encoding instead of the orthodox binary encoding; determining the group and the generations by the practical experimental statistics; and using self-adapted crossover rate and mutation rate. Finally the LTS Hausdorff distance is applied as the fitness function in the improved genetic algorithms, so that it dramatically improves the resistance to the impacts of noise and outliers in the algorithm.The author has made enough experiments using medical MRI and PET 128×128 pixel image to prove that the algorithm proposed by this paper can guarantee a good accuracy during the registration. The author also compared the algorithm proposed in this paper and the algorithms based on orthodox Hausdorff distance, and made a contrast of the differences which caused by the algorithms, after solving the same images interfered by noise and outliers. The results of the experiments proved that the algorithm is really more robust.