Automatic cephalometric landmark detection using Zernike moments and template matching

Most methodologies not only differ in the kind of techniques used to pinpoint the location of landmarks but in many other ways like: the size of training and test dataset, quality and variations covered in the dataset based on age, race, gender, etc., different studies use distinctive landmarks and different methodologies to report the results. In this Chapter, we propose a combination of local and global feature matching techniques to enhance the accuracy of the landmark localization, speed of computation and robustness to image noise. A three stage landmark strategy is used. In the first stage, the approximate landmark locations are identified using a reliable and global shape matching technique using Zernike moments and Procrustes analysis for shape alignment. Four expectation windows for each landmark are derived using these landmark positions. To bring robustness to noise only a subset of Zernike moments is used. It can be seen as a fast way of comparing two patterns in, which the finer details have been smoothed out. The Zernike moments with order five are used to find the 5% most similar images in the training set to the test image in our method. Many landmarks have clear definitions when they are searched in a small window. The RP and CP based methods can efficiently reduce the feature space and has the rotation invariance property. This leads to reduction of execution time significantly. In the second stage, the system applies rotation invariant template matching using a combination of RPs and CPs on the four expectation windows to get candidate landmark locations very close to the exact location of the landmark. Once candidate landmark points are found, the final refinement of the landmark location is done in stage three of the algorithm with slower, but more robust traditional template matching using SSD and NCC. At this stage multiple, templates are used to achieve robust search of objects in presence of different scales, object deformations, illumination variation, etc. Four templates are used to cover the variations, and the