Welcome to Pengyu Hong's Homepage

Relevant Feedback Content Based Image Retrieval Using Support Vector Machines

Content-Based Image Retrieval using relevance feedback technique allows the user to retrieve images interactively. The user firstly submits a coarse query to the system. The system searches its database and returns some images. Then, the user selects some images from those returned by the system as relevant feedbacks to refine the query. In this way, the system tries to automatically capture the high level concept and perception subjectivity of the user.

In our approach, we utilize both the positive and negative feedbacks for image retrieval. Basically, the user not only selects relevant images as the positive feedbacks but also select irrelevant images as negative feedbacks. Support Vector Machines are applied to learning the positive and negative examples. The learning results are used to automatically update the preference weights for the relevant images and retrieve new results. This approach not only better captures the information need of the user, but also releases the user from manually providing the preference weights for the positive examples.  

Discussions:

However, there is a problem. Like most image retrieval approaches, we used the global image features, which are the mixtures of the local image features. The disadvantage is obvious. An object usually consists of several primitives (e.g. regions, edges, etc) among which various contextual relations are defined. If only the global features of the object is used to represent the object, the local features of the object primitives will be mixed together. This will result in misclassification of different objects, which share similar global features while have quite different local features and contextual information. Moreover, the objects are always placed in different backgrounds. Those backgrounds provide distracting information.

Our latest research addresses the above problem.

Publication:

Update Relevant Image Weights for Content-Based Image Retrieval Using Support Vector Machines.  Qi Tian, Pengyu Hong, Thomas S. Huang, IEEE International Conference on Multimedia and Expo (ICME'2000), Hilton New York & Towers, New York, NY, July 30 - Aug. 2, 2000. 

Incorporate Support Vector Machines to Content-Based Image Retrieval with Relevant Feedback.  Pengyu Hong, Qi Tian, Thomas S. Huang. 7th IEEE International Conference on Image Processing (ICIP00), Vancouver, Canada, Sep. 10-13, 2000.