Learning visual words for content based image retrieval

This paper presents a novel framework for content based image retrieval (CBIR). Local tetra patterns (LTrP) followed by bag of visual words (BoW) is used for the feature extraction and artificial neural network (ANN) is used for index matching and image retrieval task. Existing LTrP encodes the relationship between the sample pixel and its neighbours and BoW computes the frequency of visual words present in the image. In this paper, we have integrated the advantages of both LTrP and BoW. Initially, we have detected interest points using speed up robust feature (SURF) and further local features (using LTrP) are extracted from local patch around each interest point. After feature extraction, we used BoW to obtain global representation of an image. Further, artificial neural network (ANN) is used for index matching and image retrieval task. Performance evaluation of proposed system has been carried out using average retrieval precision (ARP), average retrieval recall (APR) and F-score on two state-of-the-art databases viz. Caltech256 and GHIM10K. Performance of proposed system is compared with existing feature descriptor and CBIR systems. Performance analysis shows that proposed system outperforms the existing methods and traditional CBIR framework i.e use of similarity measurement.