Research And Application Of Feature-based Image Segmentation And Matching

Image.segmentation and matching is two important aspects of digital image processing,and in recently years,with development of information technology and reduction of hardware cost of computer,they are used extensively in many fields.In medical imaging,segmentation of patients' tumors is helpful to disease diagnosis and treatment,matching of nonomodel images from the same patient but different time, can be used to study the progress of disease and by matching of multimodel images from the same patient,doctors can get more information about the patient from only one image.In computer vision,matching can be used to object recognition and motion tracking.Based on feature regions of images,we propose a fast and effective segmentation algorithm.The LoG operator is used to extract feature regions firstly, and then,the region to be segmented will be ray-casted in larger scale.For every ray, 7^th polynomial will be used to fit gray along it.Two minima of fitting curve nearest the center point will be searched and refined as the final edge points of profile. Extensive comparison experiment shows that our algorithm outperforms methods which are based on image-gradient.Based on feature points of images,we also propose(LEAC) Loop Edge-Angle Code model which has been used for point pattern matching.Similarity LEAC describes similarity of two points' local space structures and can be used to evaluate local similarity transformation.Similarity of two points is measured by length of their similarity LEAC.Structure matching will be done with similarity of points' space structures firstly,and then local similarity transformations and clustering will be used to remove falsely matched points.Similarity condition of LEACs and clustering condition are relaxed for robustness against partially affine transformation and view change.Effectiveness of LEAC model is verified by extensive training,and extensive comparisons have shown that our matching algorithm outperform typical methods which are based on image-gradient and mutual information.We also construct a system which can be used for object recognition,image spelling,fusion of monomodel or multimodel medical images,and introduce it with different kinds of instances in details.