Image representation using the self-organizing map

Abstract

This paper introduces a new approach to image representation for multimedia databases based on the Self-Organizing Map (SOM) neural network. The distance between each image from a database and the SOM weight vectors trained on the same database is used as a representation for the image. In order to assess the performance of this proposal we compare it with a reference technique in image representation: the Thumbnails method. The results are satisfactory for an initial experiment since it was possible to identify the effectiveness of the SOM-based proposed representation. In order to verify the efficiency of the representations, a classification experiment is performed using the k-NN algorithm. For all image representation experiments, the SOM approach outperforms the Thumbnails reference technique. For example, in one experiment the representation results in a reduction of image size to 2% of its original size and the correct classification rates achieved are 83.33% and 35.42% for SOM and Thumbnails respectively. © 2013 Springer-Verlag.