A relevance feedback based image retrieval approach for improved performance

We present a new content-based visual information retrieval system that applies different existing techniques in a novel way. We use multi-scale local binary patterns to extract a set of features histogram of an image. Then we integrate it with a superpixel-based saliency model to assign weights to the features. Finally we train a group of constrained similarity boundaries using SVM to exploit the perceptual similarity between images to improve the retrieval performance. Experimental results show that the recall of our system was considerably enhanced by combining just the LBP features and the similarity boundaries and by discarding the superpixel saliency map.