Comparative Analysis of Relevance Feedback Techniques for Interactive Content-Based Image Retrieval

In recent years, with the ever-growing availability of images and videos online, Content-Based Image and Video Retrieval techniques have gained significant traction. Although state-of-the-art image and video retrieval systems perform well, there remains a need to further align these systems with users' information needs. Relevance Feedback, a process where the system refines search results based on user feedback about the relevance of initial results, is key to achieving this alignment. This thesis investigates and compares four relevance feedback algorithms used in interactive video retrieval systems, with the aim of observing how the performance of each algorithm changes based on the type of feedback and identifying the most effective approach for integration into the VISIONE system, developed by the AIMH laboratory at ISTI-CNR in Pisa. The algorithms analyzed include: the Rocchio algorithm, a classic method for query point modification in information retrieval; PicHunter, which applies Bayesian theory to update the relevance probability of images; Polyadic Search, adapted from efficient multiple-query searches; and Support Vector Machines Active Learning, which combines active learning with SVMs to enhance performance with limited user annotations. We modified the algorithms to ensure comparability, make them flexible for use with different types of feedback, and to improve their performance. Following extensive experiments aimed at finding the best version of each algorithm, a new user-friendly interface was developed. This interface facilitates qualitative testing of the studied algorithms.