Clustering of Retrieved Images by Integrating Perceptual Signal Features within Keyword-Based Image Search Engines

Abstract

Most of the Image Search Engines suffer from the lack of comprehensive image model capturing semantic richness and the conveyed signal information. Instead, they rely on the text information that is associated with the images like their names, surrounding text, etc. As a consequence, the retrieval results may return large amounts of junk images which are irrelevant to the given query. To remedy such shortcomings, we propose to enhance the performance of the text-based Image Search Engines by developing a framework that is tightly-coupling the image Semantic-Signal information for Clustering the Retrieved Images “SCRI”. Our clustering method does not rely on hard-toobtain similarity matrices of individual modalities. Instead, easily computable high-level characterization of the perceptual signal features (i.e. colorred,..., texturebumpy,.. and shapepandurate,..) are used to perform more userfriendly and intuitive searching method aiming to cluster the retrieved images based on their Symbolic-Signal information. SCRI performs partitioning, on the retrieved images, into multiple “symbolic” similar clusters in order to filter out the relevant/irrelevant images. Therefore, for images retrieved by the queryimages, SCRI performs a three-layer fuzzy filter on the symbolic characterizations, which represent the signal features, in order to achieve more accurate characterization of the diversity of visual similarities between the retrieved images. Experiments on diverse queries on Google Images have shown that SCRI can filter out the junk images effectively.