Degree Name

Doctor of Philosophy


School of Electrical, Computer and Telecommunications Engineering - Faculty of Engineering


Searching multimedia databases using features extracted from the content is currently an active research area. This thesis presents novel feature extraction approaches for content-based image retrieval when the query image is a hand-drawn black and white sketch. To facilitate robust man-machine interfaces, we accept query images with no color and texture attributes. Special attention is given to the scale and rotation invariance properties since the query and database images may vary in size and rotation angle. Several applicable techniques within the literature are studied for these conditions. The goal is to present the user with a subset of images that are more similar to the sketched query. New affine transform invariant feature extraction techniques are proposed to improve retrieval performance, and reduce the extraction and search times. The techniques are tested both generally for multi-component images and particularly for isolated shapes. The solutions are discussed for each specific application. Finally, signature-based document retrieval, which explores document retrieval from databases using human signatures, is investigated on an individual basis. Two different approaches based on spatial distribution of edge pixels are proposed for general sketch-based image retrieval. Here, the database images consist of multiple complex objects within an inhomogeneous background. One of the methods is an improved version of another, which increases retrieval performance. Both techniques exhibit scale invariance property resulting from size normalization. Rotation invariance property is achieved by employing the magnitude of the Fourier transform coefficients in two different ways. A database including 4000 artwork and photograph images is used for the experiments. One of the most important aspects of the proposed methods is that the image segmentation is not needed. This significantly improves the feature extraction process and enables the methods to be used for other computer vision applications. Edge image matching is studied as an example. Sketch-based shape retrieval, which deals with images containing an isolated shape, is studied next. This is to emphasize on the existing differences between general images and isolated shapes for sketch-based retrieval. A new approach is proposed for this task, which outperforms alternative approaches. In this approach a chain code differentiation of contour shapes is applied for contour poligonization. The geometric properties of the resulting polygon are used to extract hybrid and efficient features. A signature region extraction method is proposed as a preprocessing stage in signature based document retrieval. Several feature extraction techniques are adapted and examined for this application and the results are discussed. In all the experiments, the Average Normalized Modified Retrieval Rank (ANMRR), which was developed in the MPEG-7 standardization process, is used as the retrieval performance criterion. Feature extraction and search times are computed for time based comparisons. The proposed methods exhibit significant improvements in retrieval accuracy. There is also marked improvement in feature extraction and search speed for the proposed methods.



Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.