Learning-Based Detection, Segmentation, and Matching of Objects

Object learning is an important problem in machine vision with direct implications on the ability of a computer to understand an image. Usually, an object is defined by its appearance (the pattern of gray/color values in the object of interest and its immediate neighborhood), shape, and sometimes, by its relationships to other objects in the scene. This dissertation presents appearance-based as well as shape-based methods for object learning and retrieval. The object appearance is modeled as a Markov chain that maximizes the discrimination (Kullback distance) between positive and negative examples in a training set. The learned appearance model can be used for object detection: given an arbitrary black and white image, decide if the object is present in the image and find its locations) and size(s). Two applications will be discussed in detail: human face detection in Black and white images and heart ventricle localization in MR images. We have also developed a fully automated shape learning method which is based on clustering a set of training shapes in the original shape space defined by the coordinates of the contour points and performing a Procrustes analysis on each cluster to obtain cluster prototypes (average objects) and statistical information about intra-cluster shape variation. The main difference from previously reported methods is that the training set is first automatically clustered and those shapes considered to be outliers are discarded. In this way, the cluster prototypes are not distorted by outlier shapes. The second difference is in the manner in which registered sets of points are extracted from each shape contour. We have proposed a flexible point matching technique that takes into account both pose/scale differences as well as non-linear shape differences between a pair of objects. The matching method is independent of the initial relative position/scale of the two objects and does not require any manually tuned parameters. Our shape learning method has been used to develop a state-of-the-art hand shape-based personal identity verification system, a shape warping-based system for segmenting the Corpus Callosum in MR images of the brain, as well as an automatic system for predicting dyslexia based on the shape of the Corpus Callosum.