Combining hybrid information descriptors and DCT for improved CBIR performance

Content Based Image Retrieval (CBIR) aims to retrieves images in the database that are similar to a query image based on the contents of the image rather than metadata. The algorithm used to extract features from images is one of the most influential factors towards a CBIR system's performance. In this paper, we take a look at hybrid information descriptors (HID) as the feature extraction algorithm for our CBIR system and supplement HID with information in the compressed domain using discrete cosine transform (DCT). The HID+DCT algorithm proposed was compared with the HID algorithm on the Corel Dataset. We found out that the HID+DCT algorithm performs better than HID algorithm. We have used and compared Manhattan Distance and Euclidian Distance as distance metrics during the process of feature matching and observed that Manhattan Distance gave the best precision value for HID+DCT feature. However, the use of DCT results in a larger feature vector size which could potentially lead to slow queries. We consider using minimal-redundancy-maximal-relevance criterion (mRMR) for feature selection to reduce the size of feature vector to avoid speed related issues. We observe that the difference in precision for a feature vector reduced to almost the same size as HID's feature vector and HID+DCT is negligible.