Explorations on a neurologically plausible model of image object classification

Sourabh Daptardar
Supervisor: Prof. Amitabha Mukerjee

Abstract: In this thesis we study the problem of classification of image objects into semantic categories. Humans can perform this effortlessly, but it remains one of the most challenging problems for machines. We explore both supervised (given images and their class labels) and unsupervised (given only images) classification. We investigate the model of object recognition given by Serre et al., (2005), which has been inspired by the cortical areas in the brain and tested only on supervised image classification. We explore the variants of this algorithm and show that instead of random patches, patches at corners as well as centroids of patches also work well. Further, we show that several input classes can be combined to construct a generic category-specific filter for a single category(cars). We then propose an unsupervised classification algorithm, using normalized graph-cuts, for clustering images based on similarity. The latter method holds promise for learning categories, and also for presenting image search results, say from a search engine.