Nonlinear relevance feedback: improving the performance of content-based retrieval systems

Experimental results indicates that the proposed method yields better performance compared to a linear relevance feedback mechanism.

Abstract

A nonlinear relevance feedback mechanism is proposed for increasing the performance and the reliability of content based retrieval systems. In particular, the human is considered as part of the retrieval process in an interactive framework, who evaluates the results provided by the system so that the system automatically updates its performance based on the users' feedback. An adaptively trained neural network architecture is used for implementing the nonlinear feedback. The weight adaptation is performed in such a way that the network output satisfies the users' selection as much as possible, while simultaneously providing a minimal degradation over all previous data. Experimental results indicates that the proposed method yields better performance compared to a linear relevance feedback mechanism.