RECOGNIZING OBJECT CATEGORIES IN NATURAL SCENES USING ROI SEGMENTATION IN SVM WITH SPATIO-GEOMETRIC CONSTRAINTS AND VISUAL SIMILARITY

The image retrieval performance in Scalable databases, may be enhanced by certain paradigms like generating large Visual Vocabularies, Compact image representations and use of geometric information. This paper first reviews the above approaches and then brings out the importance of using Saliency and Spatial information, to detect the Regions of Interest (ROI). The Spatial information of the visual words in the images has not been considered in the Bag-of-Words (BOW) model. This work includes the Spatial information of visual words in the images. The Visual vocabulary is constructed using K-means. The kd-tree is used for fast vector quantisation. The image representations used are: Spatial Histogram and PHOW. The low-level features i.e., Color, Shape and Contrast information, are combined using VLAD compact image representation. The combined features are then given to Linear SVM classifier, and implemented using MATLAB.