Learning pseudo metrics for semantic image clustering

While people cluster images in terms of semantics, computers can not do too much on this job due to the use of lower-level features, which is the common way to computer workers to deal with the image recognition. To measure the closeness between feature vectors, similarity metrics play a key role and they directly impact the clustering performance. This paper develops a framework of learning pseudo metrics (LPM) for semantic image clustering practice. Multilayer perceptron (MLP) is employed to model the LPM with a set of criteria on evaluating the LPM quality and availability. Using a standard k-mean clustering technique, a comparative study is carried out to demonstrate the significance of our proposed LPM in semantic image clustering. Experiments show that the LPM-based similarity metric can produce better clustering results in terms of both impurity and robustness.