There is an increasing need for developing efficient and practical content based image retrieval techniques which can be used to organize digital libraries. In this report, we study related work in this area and identify the shortcomings of these techniques.

We propose a graph theoretic technique to define the similarity between two images. Images in the database are represented as graphs based on the semantic connectivity of the regions in the corresponding segmented images. A measure for similarity between images is considered based on the framework of ``Maximal Common Subgraphs'' which identify similar components of images.

The theoretical advantage of the technique is that the system is robust to over-segmentation, which most of the earlier approaches fail to solve. An online demo of the system is built over an image database of around $2656$ general purpose images. We have demonstrated that the proposed system is robust to image transformations like rotation, reflection, spread, sharpen and pixelize. Also, the proposed system works much better than earlier methods.