Content-based image retrieval (CBIR) is a highly pursued research topic among the database and image processing areas. The most promising techniques being developed nowadays use image processing algorithms to automatically extract visual features from each image, to represent and generate the index structures needed to search in large image data sets. However, performing simple queries over the extracted features, without the help of semantic descriptions of the images performed by human analysts, leads to retrieve too many answers that not comply with the users' needs. To circumvent this problem, researchers are studying Relevance Feedback (RF) techniques as a way to improve the interestingness of the answers. In fact, good results have been obtained in specific applications, but there are no conclusive results on the general applicability of those techniques over a wide range of application domains. In particular, there is no effective technique allowing to provide support a universal way to derive a model allowing to correct the answer of future queries based on the relevance information provided by the user over a specific query. This project aims at exploring the development of RF techniques that can be embodied into the Relational Database Management Systems (DBMS), so they can be useful to refine future queries even when the application program has not been explicitly constructed to ask for the refinement. Taking care of RF techniques in a global, server-centered approach reduces the complexity of the application development, and provides a consistency in the answers obtained through any of the client applications accessing the DBMS.
News published in Agência FAPESP Newsletter about the scholarship: