MS

 

Similarity-Based Similarity-Ba sed Annotation Propagation  IPAL MSc by Research Project

Expected Commencement: Apr 2006

Motivation With the advances in image capturing devices (e.g. digital cameras, camera phones, medical imaging devices such as X-Ray, MRI, CT etc), voluminous amounts of images are growing everyday. Current Content-Based Image Retrieval (CBIR) systems rely on feature-based similarity  between query and database images to rank and return images of decreasing similarities. Apart from the bootstrapping problem of the Query-By-Example (QBE) paradigm, the “semantic gap”  between the low-level signal-based image representation and the high-level concept-based query expectation has limited the wide spread usage of CBIR systems. On the other hand, keyword-based image search systems (e.g. Google Image Search, medical image retrieval using patient data) is still a more common way to access images. However such Annotation-Based Retrieval (ABIR) systems either require laborious manual effort to annotate the imagesImage or utilize only the associated textual information for indexing and retrieval. Even for images where annotations are readily available from social networking (e.g. Flickr), the annotations are not reusable for another image which depicts similar content. Hence a promising research direction would be to explore a way to transfer available image annotations to new images  based on visual similarity and related contextual similarity.

Research Focus This research project will explore a novel framework to propagate existing image annotations to other images (without annotations or with partial annotations) based on the visual similarities and the contextual similarities among the images. More specifically, the student will study and solve the following research issues:   A structured image representation (beyond low-level features) based on local semantic visual concepts and their spatial relationships (e.g. hierarchical part-based object models, region based graph models)   A similarity measure appropriate for matching two structured image representations   A decision procedure to propagate annotations (e.g. keywords with frequency counts) associated with a group of images to an image (without annotation or with partial annotation)   An auxiliary representation based on contextual information (e.g. time, location, personal data etc in the case of consumer images, modality, anatomy, pathology etc in the case of medical images)   An associated similarity measure and decision procedure based on the auxiliary context representation to enhance the annotation propagation process enabled by the image-based similarity There could be different sources of known image annotation (e.g. medical atlas, medical case •

•

•

•

•

sheets, social networks etc). The student is expected to apply and evaluate the research outcome on large datasets (e.g. ImageCLEF).

 

  The candidate should possess a B.Sc. or B.Eng degree (or equivalent) in Computer Science or Computer Engineering from a recognized university. The candidate should have basic knowledge in image processing, computer vision, information retrieval and strong programming skill in C/C++/Java. The candidate must be able to communicate well in English, self-motivated, and work independently. The candidate will conduct the research work in Singapore for at least 3 months per year during the candidature. This project has sufficient scope for extension towards a PhD research topic

Contact: Dr Lim Joo Hwee Director (Singapore) French-Singapore IPAL Joint Lab Institute for Infocomm Research 21 Heng Mui Keng Terrace Singapore 119613 Tel: +65 6874 6700 Fax: +65 6775 5014 Email: [email protected]