Automatic Graph Based Clustering for Image Searching and Retrieval from Database

Content-based image retrieval and searching is one of the most burning issues in the field of multimedia computing. Human perception is not understood well enough to automate the retrieval process. In this work we have designed a system for content-based image searching. This system uses multiple cues (features) for image searching and retrieval. Since most of the features have some drawbacks, we use the cues that are free from drawbacks like geometrical transforms and viewpoint variation. We present the results based on these cues. A heuristic for combining the result of different cues to increase the accuracy of the system is developed. Databases of different size were used to estimate the accuracy of the system. Global shape descriptor of images and object based descriptors are extracted for the retrieval of images. Multimedia databases are very big in size, so we cannot go for exhaustive searching of images from these databases. For this purpose an automatic graph-based clustering algorithm is developed to reduce the searching time of the images from the database. The proposed algorithm works on the concept of minimum spanning tree that removes the inconsistent edges from tree, based on the dynamic threshold provided to the algorithm. The proposed algorithm reduces the search time for the retrieval without much loss in the accuracy. We found out that careful combination of the different cues, based on our proposed heuris­tic, can increase the retrieval accuracy up to a noticeable extent.