Research On Uncertainty Modeling And Matching Method Of Geometry Features In Image Registration Field

Feature matching is a basal issue in the fields of Computer Visual, Image Understanding and Photogrammetry and Remote Sensing that have been researched for a long time. Its aim is to find the correspondences among features that extracted from images taken at different times, from different viewpoints or by different sensors, or features extreacted from image and model. It is the key step of image registration, target recognition, target 3D reconstruction and image sequences analysis.In the feature matching problem, most exsiting methods do not consider the imprecision of the feature parameters, but take them as the true attributes of features. In addition, during the process of the finding corresponding features, people consider the features as same important. These factors may impact the accuracy of the matching result and even lead to a failure. Therefore, it is necessary to analyze and study them. The paper analyzes and models the uncertainty of feauture such as corners, line segments and so on. We take image registration for example to research the problem of feature matching that considers uncertainty. The main achievements and contributions of the dissertation are listed as following:(1) An algorithm for uncertainty modeling of features and similairy computering based on Vague Set is proposed. It firstly analyzes and models the uncertainy of position and length parameters; then defines and computes the similarity function of features which consider uncertainty. The proposed algorithm overcomes the problems of big matching error and difficult to find optimal corresponding relation which exists in conventional algorithms. The proposed algorithm considers both the agree degree and disagree degree which makes the result more rational.(2) Proposes a corner extraction algorithm that takes both gray information and edge into account and a point matching strategy that considers uncertainty. In corner extraction algorithm, it firstly extracts and groups edges; then computers the attributes and fractal signature values of edge poins; finally extracts the satisfied points. Comparing to the conventianl algorithms that only take use single information of gray or edge, the proposed algorithm is prominent in robustness, repeat rate and position accuracy. The paper brings uncertainty into the problem of point matching. It firstly abandons the unsteadiness features using robustness values; and then solves the similarity function using Vague Set theory; finally, designs appropriate strategy to find the corresponding features. When computing the parameters of the transformation, robustness values are used to weight the effects of the features which will increase the influence of robust feature and improve the registration accuracy.(3) Proposes a new line segments matching algorithm considering uncertainty. It firstly extracts line segments from image, then groups the two line segments into a Feature-Line-Pair (MLP) according to the lengths, robustness values and their distances and computes its atrributes, finally, finds the optimal solution using branch-and-bound method. Comparing to other features, MLP has some advantages, such as with plenteous information, can be easily extracted and can extract control points easily. By limiting the angle and robustness value, the robustness of the feature is improved. The limitation of the robustness values and angle between the two lines of the MLP can improve the robustness of MLP and decrease the error caused by the orientation excursion. When computing the similarity of MLPs, we mainly use the angles which can decreasese the effection of the rupture. Comparing to the conventional methods, the proposed strategy of matching not only considers the similarity of the parameters, but also takes the spatial relation and uncertainty into account which makes the method be robust to noise and improves the registration accuracy.(4) Proposes a multi-feature matching algorithm based on feature-index. Then we take MLPs and affine invariant feature which is newly proposed in this paper according affine geometry for example to validate the algorithm. It ultilizes the idear of palette and firstly constructs base feature-index-table which makes up of index value and detail; then constructs feature-index-table in reference image; finally, matchs the features using the tables. The introduction of feature-index-table improves the efficiency of matching and makes the algorithm more suitable for the real-time tasks. The reliability and accuracy of the matching result are ensured by throwing off the unrobust features that with small robustness values. In addition, the proposed matching strategy can guarantee that the absence of one kind of features does not impact the final accuracy of the algorithm. The idear of using multi-feature to register images not only expands the usage range greatly but also improves the accuracy of registration.