Image Retrieval Using Patch Based Feature

Abstract

Image retrieval from large datasets has become an area of wide interest nowadays in many applications. In this thesis we present a content-based image retrieval system that uses patches from image as visual features to describe it. We adopted Dense Micro-Block Difference as local feature in our CBIR system. The used features are based on idea that small patches in a texture image exhibit a characteristic structure and, if captured efficiently, discriminative information can be obtained. Such features are encoded using bag of visual words to obtain an image descriptor which considers higher order statistics. In this step, first, the local features are extracted from the training images and then exemplar features are chosen as the textons (using K-means clustering). These textons are used to label all the features from training and test images. Our experimental evaluation of the system is based on different image datasets. From the experimental results, we found out that DMD significantly outperforms other features.