Loading...

Image Registration Using Graph-based Methods

Taheri Dezaki, Fatemeh | 2015

2117 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 47756 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fatemizadeh, Emadeddin
  7. Abstract:
  8. Nowadays, image registration is considered as one of usual issues in medical researches whose new findings are expanding outstandingly and it has reached a high level of maturity. Generally speaking, image registration is a task to reliably estimate the geometric transformation such that two images can be precisely aligned. With respect to different uses of image registration in medical applications, it has attracted the attention of many scholars and there has been made significant improvement in this realm. Image registration is still one of the active branches in medical image processing due to its wide applications and problems. Graphs, thanks to their geometric structures and intuitive concepts, are powerful and universal data structures able to explicitly model networks of relationships between substructures. These features provide an appropriate base for studying other problems in the state of their corresponding problems in graph theory. The vast majority of processing approaches rely on object representations given in terms of feature vectors which cause the effect of losing contained structural information. Recently, however, a growing interest in graph-based object representation can be observed. Keeping structural information is critical in some applications such as image registration. In order to achieve this goal, there is a need for a different representation space and appropriate analysis methods. In most practical fields, structural information can be represented by graphs. The process of evaluating the similarity of two graphs is commonly referred to as graph matching. Applying this procedure in image registration, which needs an efficient tool for matching extracted feature correspondences, is a new robust approach. According to the fact that the quality of image registration’s performance is directly related to the results of matching step, developing a robust method in the presence of noises and outlier points plays an important role. Using the pairwise agreement as a second order interaction is a key point in exploitation of structural properties. It is unlike the conventional methods which just the use point-wise similarity as the first order interaction. In this thesis, we surveyed different existing graph matching methods and with regard to the weakness and challenges of previous matching algorithms, we proposed two approaches with the use of graph concepts to improve the robustness and accuracy of matching correspondence points. The suggested algorithms make the process of matching less sensitive to the presence of outliers existing among landmarks and as a result, it provides a better performance in the image registration process. In the first approach, we consider a feedback loop to modify the results with respect to the previous ones in such a manner that it diminishes the effects of outliers and yields a way of detecting them. The second approach was designed in a way that probable correspondences have more weight in the cost function; in fact, there is more attention to inlier points than outliers. The idea of penalizing cost function in conflict assignments was proposed to enforce the algorithm to a better response. As the modifying effect of each proposed methods, they can combine in one process of matching, which results in significant improvements. Experimental results on synthetic and real databases indicate the effectiveness of the proposed method in terms of both robustness and accuracy compared with several conventional graph matching methods
  9. Keywords:
  10. Image Registration ; Quadratic Assignment Problem ; Graph Matching ; Pairwise Affinity ; Correspondence Detection ; Outlier Robust Matching ; Sparsity Constraint

 Digital Object List

 Bookmark

...see more