A method for dynamic indexing of large image databases

The fascinating phenomenon of optical holography has been studied for nearly three decades. A new associated memory model based on the mathematical representation of optical holography was proposed in 1990. This approach was subsequently adapted to the problem of indexing large image archives. For this, trained associative memories based on the holographic model were used. The training process involved several iterations over the stimulus-response pattern associations during the encoding of the memory. This method, however, has severe limitations when it comes to dynamic indexing of images because of the need for retraining the associative memory whenever additions or deletions of images are made in the database. In this paper we describe an encoding and decoding scheme for indexes of image archives based on a holographic model that does not involve training. As there is no training overhead, this method is suitable for dynamic indexing of image databases that are subject to frequent additions and deletions. We investigate the performance of this scheme using simulations and compare it with the method based on the trained holographic memory model. We also report the results obtained with real image data. The untrained memory model with the retrieval method developed by us compares well with previous results for the effective operational range of the holographic memory. More importantly, the untrained memory gives robust performance when a large number of additions and deletions are made in the database.