ABSTRACT
This work introduces PicHunter, an image retrieval system that uses a unique method for relevance feedback in which the system considers the full history of user selections when estimating the user’s objective picture. PicHunter does this by using Bayesian learning, which is based on a probabilistic model of a user’s behaviour. The model’s predictions are merged with the search results to determine the likelihood associated with each image. The probabilities are then utilised to choose pictures for presentation. An offline learning method was used to fine-tune the details of our user behaviour model.
To be clear, our research used the simplest conceivable user interface, but the method may be simply integrated into systems that enable complicated queries, including the majority of previously presented solutions. Despite this limitation and minimal picture attributes, PicHunter is able to detect randomly picked targets in a library of 4522 photos after presenting an average of just 55 groups of four images, outperforming chance by more than tenfold. We anticipate that the strategies discussed here will improve the performance of current picture database retrieval systems.
Authors: Ingemar J. Cox, Matt Miller, S. M. Omohundro, Peter Yianilos | NEC Research Institute, Inc., Princeton, New Jersey, USA.
Publisher: Proceedings of 13th International Conference on Pattern Recognition (IEEE Xplore)
Citation: I. J. Cox, M. L. Miller, S. M. Omohundro and P. N. Yianilos, “PicHunter: Bayesian relevance feedback for image retrieval,” Proceedings of 13th International Conference on Pattern Recognition, Vienna, Austria, 1996, pp. 361-369 vol.3, doi: 10.1109/ICPR.1996.546971.
Follow JSTOR ONLINE for more updates!