Partial duplicate image retrieval using fast visual word generation technique

Abstract

Content-Based Image Retrieval (CBIR), a technique which uses visual contents to search images from large scale image databases according to users interests. In most of image retrieval systems, accurate and partial relevant matching is main issue. Partial duplicate image retrieval is most complicated task as region of interest may undergo occlusion, rotation, scaling and other transformations. Retrieval of these partial duplicate images is done by exploiting spatial information of image features. In this, position and orientation of visual word will be used to generate a COP coordinate which will be used for matching. COP coordinate are robust to rotation, scaling so better matching with high precision may be obtained. Bag of Visual Words (BOW) approach is widely used in most of Image Retrieval systems and for such systems, assignment of features to their appropriate visual words is more time-consuming. Weighted tree indexing technique is based on probabilistic tree model which uses co-occurrence of visual word within local region. This method has reduced number of backtracking while maintaining the accuracy. This system uses bag of visual words approach with weighted tree technique to generate visual words and Combined Orientation and Position (COP) graph to find more accurate matches to increase accuracy of retrieved images. Experimental results on various datasets shows that mean average precision of system is increased, leading to better retrieval accuracy.