Improving the quality of K-NN graphs through vector sparsification: application to image databases
International Journal of Multimedia Information Retrieval
The Van Houten library has removed some of the personal information and all signatures from the approval page and biographical sketches of theses and dissertations in order to protect the identity of NJIT graduates and faculty. ABSTRACT LOCAL SELECTION OF FEATURES AND ITS APPLICATIONS TO IMAGE SEARCH AND ANNOTATION by Jichao Sun In multimedia applications, direct representations of data objects typically involve hundreds or thousands of features. Given a query object, the similarity between the
... query object and a database object can be computed as the distance between their feature vectors. The neighborhood of the query object consists of those database objects that are close to the query object. The semantic quality of the neighborhood, which can be measured as the proportion of neighboring objects that share the same class label as the query object, is crucial for many applications, such as content-based image retrieval and automated image annotation. However, due to the existence of noisy or irrelevant features, errors introduced into similarity measurements are detrimental to the neighborhood quality of data objects. One way to alleviate the negative impact of noisy features is to use feature selection techniques in data preprocessing. From the original vector space, feature selection techniques select a subset of features, which can be used subsequently in supervised or unsupervised learning algorithms for better performance. However, their performance on improving the quality of data neighborhoods is rarely evaluated in the literature. In addition, most traditional feature selection techniques are global, in the sense that they compute a single set of features across the entire database. As a consequence, the possibility that the feature importance may vary across different data objects or classes of objects is neglected. To compute a better neighborhood structure for objects in high-dimensional feature spaces, this dissertation proposes several techniques for selecting features that are important to the local neighborhood of individual objects. These techniques are then applied to image applications such as content-based image retrieval and image label propagation. Firstly, an iterative K-NN graph construction method for image databases is proposed. A local variant of the Laplacian Score is designed for the selection of features for individual images. Noisy features are detected and sparsified iteratively from the original standardized feature vectors. This technique is incorporated into an approximate K-NN graph construction method so as to improve the semantic quality of the graph. Secondly, in a content-based image retrieval system, a generalized version of the Laplacian Score is used to compute different feature subspaces for images in the database. For online search, a query image is ranked in the feature spaces of database images. Those database images for which the query image is ranked highly are selected as the query results. Finally, a supervised method for the local selection of image features is proposed, for refining the similarity graph used in an image label propagation framework. By using only the selected features to compute the edges leading from labeled image nodes to unlabeled image nodes, better annotation accuracy can be achieved. Experimental results on several datasets are provided in this dissertation, to demonstrate the effectiveness of the proposed techniques for the local selection of features, and for the image applications under consideration.