An Interactive Content-Based Image Retrieval System

Content-Based Image Retrieval (CBIR) systems have attracted large amounts of research attention since 1990's. Contrary to the early systems, focused on full-automatic strategies, recent approaches introduce human-computer interaction into CBIR. RETIN is the on-line image search system developed in the ETIS lab (ENSEA, France). A web version of the software is available (beta version)!

• one click to select positive examples,
• two clicks to select negative examples,
• then update the search engine.

We believe that combining rich image description with learning techniques may provide good solutions for image retrieval tasks. We focused on user interaction systems because it is a natural way to get examples for learning (what the user is looking for). Our RETIN system adapts machine learning techniques and statistical modeling to CBIR. Relevance feedback is modeled as a binary classification. Kernel functions and kernel-based reduction and classification techniques are used. RETIN is designed to grasp a user's query concept quickly despite time and sample constraints. Active learning framework is proposed to optimize this user interaction. See our publications for detailed explanations.

• Contact: Matthieu Cord (ETIS, ENSEA)
• RETIN Powerpoint animation  
• For more background information, see Matthieu Cord's Istanbul presentation