Adaptable text and image retrieval systems using relevance feedback

The purpose of this work is to develop adaptable and robust search and retrieval systems for both text and image databases. The proposed systems referred to as "model-reference text retrieval system (MRTRS)" and "model-reference image retrieval system (MRIRS)" are inspired from the well-known theory of model-reference adaptive control systems. A learning methodology that incorporates users' information and expertise via relevance feedback to improve the relevancy of solutions is presented. This methodology relies on a limited number of single-term as well as multi-term queries for text databases and on a limited number of training samples for image databases.; The learning in MRTRS involves three phases that are: (i) initial model-reference learning, (ii) model-reference following, and (iii) relevance feedback learning from expert users. The initial model-reference learning involves capturing the behavior of a reference text retrieval model, when this is available, or simply the results of an indexing system. The model-reference following is implemented in dynamic conditions, when documents are added, deleted or updated. New relevance feedback learning methods are developed for single-term and multi-term queries from multiple users using either score-based or click-through selection feedback.; We propose two different MRIRS retrieval systems that can operate in an online fashion or in a batch mode. The first retrieval system uses several regulators working independently from each other, though they are influenced by the users' feedback. Each regulator transforms the original query image into a mapped version with the goal of driving the error signal between the output of the retrieval system and that of the expert users to zero. The second system uses a single regulator to deliver a single mapped version of the submitted query image. Although, structurally more complex, the multiple regulator image retrieval system involves lesser number of parameters to fine-tune in response to either model reference or relevance feedback learning.