Image annotation is an integral and important task for image retrieval. Automatic image annotation has been studied for quite some time now, but there is still enough scope for improvement considering the challenges associated with it. Existing systems focus on reducing the semantic gap between image and text using various heuristic, probabilistic or learning based approaches. Often, the automatic annotation tools segregate an image into discrete objects and try to annotate them. In doing so, they run the risk of missing out on important information conveyed by the image as a unit (concept). In this paper we propose a novel two-pronged probabilistic model based on a concept graph build out of such concepts which not only helps describe the objects in the image but also captures the essence of the image. To this end, we also employ an established community detection algorithm over the concept graph to identify the closest possible annotation for the image. A rigorous set of experiments on a standard dataset substantiates our proposed model's efficiency and efficacy.