Generation of Visual Patterns from BoVW for Image Retrieval using modified Similarity Score Fusion

The Bag of Visual Words (BoVW) turns up to be an efficient method to represent images for Content Based Image Retrieval (CBIR). Despite their significant usage, the traditional BoVW method has low discriminative power and fails to provide spatial information, which increases the false positive images and reduces the precision values. To address the first issue, a novel way of identifying a set of visual words unique for each category, named as Visual Patterns (VP) is proposed. Also, the weight for the respective VPs and a new way of score calculations for similarity matching with the database images are proposed. Then, to address the second issue of enhancing the spatial information, late fusion of Gabor filter features along with VP is proposed. As a consequence, VP provides better discriminative power and Gabor filtering, taking advantage of its complementary clue, provides spatial information. Hence, it helps to reduce the false matches and improves the precision values. Experiments are carried out on the popular datasets, namely, Caltech 256, Oxford 5K and Inria Holidays datasets along with Flickr 1M dataset. The proposed method is compared with other BoVW based models and proved that the MAP value is improved 0.50 times from the basic BoVW model.