An Adaptive Similarity Measure for Browsing Image Databases

Engines for browsing image databases are usually based on predefined features for selecting the images which are more similar to the one clicked by the user. However, users' feedbacks and search patterns can be used to gather a set of examples to adapt the feature extraction process in order to reflect the similarity criterion which guides the users' selections. The proposed approach uses a graph-based image representation that denotes the relationships among regions in the image and a recur-sive neural network which can be trained to extract an optimal feature vector from the graph-based representation. The graph-based representation combines both structural and sub-symbolic features of the image, while the neural network can be trained to map graphs extracted from similar images onto near points in the vectorial feature space. We report a set of preliminary experiments on real world images.