|Giuseppe Passino's Homepage|
Since October 2006 I am a research student in the Multimedia and Vision (MMV) group of the Electronic Engineering department at Queen Mary, University of London (QMUL), UK. I am supervised by Prof. Ebroul Izquierdo.
My research topic is related to the Content Based Image Retrieval (CBIR), and in a glance I'm working on an image analysis technique that is able to perform the image classification based on the extraction of image parts (patches) and on the graph-based analysis of their interrelationship.
The Content-based Image retrieval is, according to Wikipedia, "the application of computer vision to the image retrieval problem, that is, the problem of searching for digital images in large databases". This means that the retrieval of an image is performed by the means of the actual image content rather than on manual tagging or on text information related to the image. The problem has being given a great attention by the academic and industrial scientific community and it is very relevant nowadays, because of the large range of possible applications, related to the accessibility and usability of huge amounts of graphical data - the content is not usable and not useful at all if it is not easily accessible. The applications of such systems go from the application on the Internet (the so called semantic web), to large databases of images in home (e.g. photo albums) or commercial system (e.g., security camera shots, public images databases, and so on).
The analysis of an image can be performed at different abstraction layers: the images can be grouped according to some low level feature (e.g., all the images whose predominant colour is blue) or by semantic content (e.g., all the images representing "the sea"). The latter approach is of course more challenging, but it requires the possession of additional semantic knowledge (e.g., the sea is blue) that is not included in the processed images. These two abstraction levels lead to two main different extraction approaches: the query by example (where the system has to find the most similar images to a given example query image), and the semantic query (where the desired outcomes should have a semantic content related to that one specified in the query, likely in a textual way).
Among the systems based on the exploit of the semantic content of the images, a particularly promising approach is the so called human (biological) system, that, in a similar fashion as the humans do, analyzes the objects in an image by the inspection of the object parts. The resulting - commonly referred as part based - systems have to address many issues, as the exploiting of the semantic knowledge, the identification of "suitable" object parts and the management of the inherent complexity of the process. These issues have not been optimally answered yet - the proof of this fact is that the outstanding systems for the part-based approaches (the bag-of-features approach) rely on classification paradigms recovered from other recognition fields (like the semantic text classification), that ignore some kind of important information source - for example the spatial relationship between basic object constituents.
The system that I am currently working on tries to extract semantically meaningful object parts (macroscopic parts) through local operators, and to analyze their interrelationship, inferring the presence of a particular object from the analysis of the properties of the single parts as well as the parts mutual position. The correct addressing of the interdependence analysis should lead to outperforming results due to the full employment of semantically rich information.
|Last update: 23/09/2007|