System Design of a Super-Peer Network for Content-Based Image Retrieval

Super-peer networks inherit the advantages of P2P networks, such as pooling together the shared data (images in our system) across peers, self-organizing, and fault-tolerance. In addition, they take advantage of the heterogeneity of capabilities across peers in load-balancing and network adaptation. A super-peer node operates as an equal peer, and as a server/parent to a set of peers. Content-based image retrieval (CBIR) is concerned with using the image visual content, such as color, texture, and shape, to search and compare images from large-scale image database. In this paper, we adopt a super-peer network for search and retrieval of images across a group of peers´ shared image collections. In the proposed system, an image is represented by a multidimensional feature vector, of dimension d, that is extracted from the image color and texture features. We present a novel organization of super-peers as a structured overlay P2P network (specifically CAN). A super-peer maintains a zone in the d-torus virtual space that contains a set of global image categories (clusters). In addition, a super-peer manages the set of peers that hold images belonging to those global categories. That is the proposed super-peer network is both structured and clustered according to the peers´ image collections. A semi-fuzzy clustering algorithm is used for both clustering a peer image collection and for assigning a peer image cluster to at least two parent super-peers. We illustrate that the use of a structured and clustered super-peer network guarantees the retrieval of the most similar images to a query image. Scalability analysis demonstrates that the extra storage requirement for a peer is less than 0.1% of its total shared images´ size, and less than 9% for a super-peer.