Unsupervised Image clustering

Extracting semantic information from images has attracted much attention in the domain of computer vision and image processing. Areas like face recognition, detection, tracking etc. work on identifying semantics in images. In this paper we attempt to cluster images based on their semantic content. The approach involves segmenting the image at different scales and extracting interesting patches in the image. A over sized dictionary of these patches is constructed. Every image is then back projected on this dictionary to obtain a sparse feature vector for the image. This is then used to cluster the image. If we label the clusters, a new image can also be assigned to one of the existing clusters to achieve a classification. Unsupervised clustering has wide applications in information retrieval and image search. The algorithm can be possibly used to filter image search results. The paper describe the algorithm, the various techniques used and the results obtained.