Loading...

Developing Robust Image Similarity Measure in Feature Based Image Registration

Majdi, Mohammad Sadegh | 2015

2064 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 48248 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fatemizadeh, Emad
  7. Abstract:
  8. Image registration is an important preprocessing step in analysis of medical images. Detection, Treatment plan, disease grows process analysis and assistance in surgical applications are some of medical images applications. We need to be able to compare different modalities in medical images such as X-ray, PET, MRI, and CT... , or sometimes doctors need to take images of a patient in a same modality but in different times and directions. In which in order to be able to do theses comparisons we need to first align these images by using image registration methods. Image registration is an image processing method in which tries to find a geometrical transformation that would map different images into the same dimensions. Each registration methods can be divided into three main components. Registration transform, similarity measure used in finding the correspondence points in different image, and an optimization method that would minimum the registration error. Image registration first developed to align images of brain with different modalities, in order to align these modalities one needed to find a rigid transformation based on shape and position of brain. After a while scientists developed a new type of transformation called Affine transformation that would able them to add some additional factors including scale factor and sectional slicing that would mostly being used in calibration of different scanners or grows rate difference among different people. Nowadays by the development of faster and far more advanced computers and scanners, new Non-rigid transformations has been introduced, these methods can be used to model soft tissue changes during imaging or surgery and also anatomical changes caused by organs like blood pumping by the heart. Today a vast variety of methods for different application has been developed but most of them only applies to the images in the absence of noise and outliers, or even the ones that developed to work under these conditions, cannot find the perfect transformation especially by the increase in amount of noise ad outliers. In this thesis we have introduced some novel ideas to map different images in the presence of noise and outliers. 4 different methods has been developed in which the third method is the combination of first two methods, also in the fourth due to its similarity of results to the second method we only introduced the mathematical algorithms. In the first method we have tried to add a sparse term to the optimization function of famous CPD method, and so cover the effect of noise and outliers. In the second method we have used a different statistical distributions that is more robust to noise and outliers, in order to get to such distributions, we have tested different statistical distributions mathematically and conceptually and finally went with the Laplacian distribution. In the fourth method instead of norm 1 as it is in Laplacian distribution we have used Norm p distribution which, since we solved the Laplacian distribution mathematical in a general form we were able to use those methods for the norm p distribution as well with an slightly different approach. And finally we have applied our methods on different medical and non-medical databases and achieved great results in all aspects from the accuracy of finding the correspondence to minimum square error and even number of iteration and the time needed for convergence
    in most of the tests
  9. Keywords:
  10. Image Registration ; Nonrigid Transformation ; Affine Transform ; Outliers ; Sparsity Constraint ; Nonrigid Transformation ; Point Base Registration ; Rigid Transform

 Digital Object List

 Bookmark

...see more