Presenter

Mr. Frederik Temmermans - ETRO [Email]

Abstract

Visual search is the task of identifying visually similar objects. In computer vision, and more specifically in the domain of still images, this task is typically accomplished by comparing similarity of a feature set extracted from the reference image and from the target images. These image features describe properties of the image, or identified objects in the image. Typical features include color, texture or shape properties. Visual similarity based on image feature extraction forms the basis of several applications, including searching for similar images, content-based image retrieval, recognition and classification tasks.

This research contributes to the domain of visual search at three different levels. The first level focuses on the fundaments of visual search, such as image feature extraction for classification and recognition purposes. The second level discusses interoperability, focussing on interfacing and standardization related to representation, sustainability and interoperability of image repositories. The final level covers human-machine interaction. Together, these three aspects cover the whole visual search application chain, where the interoperability part forms the bridge between fundaments and human-machine interaction. These three levels are explored in two application domains: medical and mobile.

In the medical domain, at the first level, the classification task of microcalcifications, digitized using micro-CT is discussed. Microcalcifications are tiny spots of calcium that may occur in the female breast. They often are an early sign of breast cancer. Many studies indicate that their shape is an important factor to discriminate between benign and malign abnormalities. Micro-CT allows for visualizing them at a much higher resolution than contemporary breast imaging modalities. This is the first study to adopt shape descriptors on high-resolution 3D renderings of microcalcifications, to statistically analyze their shapes and examine the feasibility of an automated shape-based classification.

The second level handles interoperable image search. Interoperable search requires unified communication protocols and metadata sustainability. In this context, contributions delivered to JPSearch are presented. JPSearch is an activity within the JPEG standardization committee. It aims at providing a standard for interoperability of still image search and retrieval systems. This thesis explains and motivates the standard and portrays how it came to be. Furthermore, it is shown how the standards have to evolve along with advances in web technologies such as the so-called “Web 2.0” and the semantic web. A proposal for defining a standardized REST interface, complementary to the current standards, is presented. The usage of the standards is illustrated with examples in the medical domain and mobile image retrieval.

The final block of the application chain is to bring the technology back to the user. Via user interfaces, clinicians interact with the underlying technology, without the need to know anything about how this underlying technology works. In this way, their efficiency can be improved in terms of accuracy, time efficiency or effective reporting. This interaction between human and machine is the final level in the application chain. An innovative application for mammography experts is presented that is designed to fit in their routine workflow and aims at improving their efficiency.

In the mobile domain, this thesis portrays the mobile application market and emphasizes the need for interoperability. Two real world content-based image retrieval applications are presented. The first application focusses on recognition of artworks while the second application handles the automated annotation of buildings, which are, in contrast to the paintings, 3D structures which can be captured from different angles. At the fundamental level, the novelty situates at the combination of local and global features and combining similarity measures respectively.

Short CV

Licentiaat in de Informatica