Optimizing feature-vector extraction algorithm from grayscale images for robust medical radiograph analysis

The principal axis projection (PAP) technique developed for robust image representation has been optimized for delicate grayscale image recognition. The PAP technique utilizes the edge information in four principal directions in an image, and generates a feature vector very well preserving the human-perception of the similarity with a great dimensionality reduction. The optimization was carried out for the algorithm in determining the edge-detection threshold, projecting edge flags onto principal axes, and smoothing vector elements. The number of templates for image recognition was also optimized utilizing the generalized Lloyd algorithm. As a result, the cephalometric landmark identification, one of the most important clinical practices in orthodontics of dentistry, was successfully carried out.