. . . . . . . . . "We further present a novel semi-supervised shape classifier for historical image databases. The retrieval model is gradually learned in a self-training fashion using only a limited number of labeled instances. We then propose a new distance measure which mitigates variances in shape complexity. Our novel observation shows that shapes exhibit different shape complexities in the time series representation. In this thesis, we introduce a novel algorithm to measure this shape complexity, and we show a technique to adjust the traditional Euclidean distance so it is invariant to shape complexity. The experimental evaluations show a significant improvement in the classification accuracy of our novel semi-supervised learning method over the traditional shape classifiers. Finally, we address some myths in Dynamic Time Warping (DTW). As one of the earliest similarity measures for time series proposed in the literature, DTW has been widely discussed in the literature. In this thesis, we address some persistent myths about it, including some that have limited its adoption."@en . "Digital collections of historical manuscripts have opened up new opportunities for the computer science community. It has been shown that advanced computing tools can help historians and genealogists better analyze historical documents. While a large number of the work in this area has been focused on historical texts, in this thesis, we develop novel classifiers that help analyze historical manuscripts of images. Many existing retrieval models on image analysis require a large number of labeled data. However, in the context of historical studies, labeled data are difficult to obtain. In this thesis, we propose two novel retrieval models for exploiting historical image databases using a limited number of labeled data, or no labeled data at all. We first show a general model for annotating images in historical archives. In this model, a weighting parameter is required to combine multiple image features. We present a novel one object classifier to learn this parameter using unlabeled data. Unlike other existing learning methods, our new one object classifier requires no prior knowledge in terms of data/class distribution. The experiments show that our techniques are able to find the appropriate weighting parameter for different historical image datasets, where the weighting parameter varies."@en . "Data mining techniques on historical image databases"@en . "Dissertations, Academic"@en .