Motivation and Goal
The goal is to retrieve images in database that are similar to the query the user has in mind. Relevance feedback is a technique to let the user interact with the system by giving examples so that the system has more information of what the user needs. The key is how to make the best out of the feedback information.
More formally, we define our problem as follows: Given the database images that the user considers similar to a query, rank all database images from most to least similar to the query. It is worth mentioning that the query does not have to physically exist, but can be a concept the user has in mind. All the system needs are the images that the user identified as similar or relevant to the query.
All the logo images in the database have been labeled with design search paths. Design search paths were assigned by trademark and patent experts. Based on design search paths, we build feature vectors, which we call high-level semantic feature vectors. They are binary vectors, each dimension being a feature representing the presence (feature value equal to 1) or absence (equal to 0) of a specific object, concept, or shape.
We present a procedure to retrieve objects based on the high-level semantic feature vectors. The assumption is, if two high-level semantic feature vectors are relevant, then they share at least one non-absent feature.
Any suggestions or comments are welcome. Please send them to David Liu.