Automatic Image Annotation Using Random Projection in a Conceptual Space Induced from Data

The main drawback of a detailed representation of visual content, whatever is its origin, is that significant features are very high dimensional. To keep the problem tractable while preserving the semantic content, a dimensionality reduction of the data is needed. We propose the Random Projection techniques to reduce the dimensionality. Even though this technique is sub-optimal with respect to Singular Value Decomposition its much lower computational cost make it more suitable for this problem and in particular when computational resources are limited such as in mobile terminals. In this paper we present the use of a 'conceptual' space, automatically induced from data, to perform automatic image annotation. Images are represented by visual features based on color and texture and arranged as histograms of visual terms and bigrams to partially preserve the spatial information [1]. Using a set of annotated images as training data, the matrix of visual features is built and dimensionality reduction is performed using the Random Projection algorithm. A new unannotated image is then projected into the dimensionally reduced space and the labels of the closest training images are assigned to the unannotated image itself. Experiments on large real collection of images showed that the approach, despite of its low computational cost, is very effective.