Image summarization using topic modelling

With the rapid advancement in technology it is easy to take pictures, and easier still to share them with the world at large. In this digital age the image data we have has grown manifold over the past few years. Digital Cameras have allowed the capturing and storing of large number of images highly inexpensive and hence we can´t go on a trip without taking several hundreds of photographs. We take repeated shots till we find the perfect one. As a result, in a collection of images we end up having recurring and similar images which we would like to omit. Manually going through such a huge collection and picking out the best images is a highly laborious task. This paper deals with devising a way of summarizing a given collection of photographs to represent a distinct set of representative images. This would save the users a lot of effort while allowing the selection of the best images of the set which also represent the entire set. Here we employ a modified Latent Dirichlet Allocation technique, a generative probabilistic model, to partition the images from a `Bag of words´ representation created using Scale Invariant Feature Transform (SIFT) vectors and then clustering these vectors into bins. We validate the results using subjective analysis based on 3 metrics by the people providing the image collection and also by a more general set of people.