Image Retrieval for Multi-image Queries Handling Hidden Classes

The image retrieval system is used for browsing, searching and retrieving images from a large database of digital images. In the proposed system, Content-Based Image Retrieval (CBIR) handles the predefined classes using low level features. To improve the accuracy of the retrieval, color and texture features of the image is extracted, which is represented as color co-occurrence matrices. In retrieval, complexity of selecting a query object in single image query is high. To avoid this problem, multi-image query is used to perform the retrieval. Support Vector Machine (SVM) is used to construct the classifier for pre-defined classes. However, in a large-scale image collection, some image classes may be unseen. These unseen image classes are termed as hidden classes. In order to handle the hidden classes, the unclassified images are clustered, based on color and texture feature using K-means clustering algorithm. The queries associated with the hidden classes cannot be accurately answered using a traditional CBIR system. To handle these hidden classes, a robust CBIR scheme is proposed that incorporates a novel query detection technique, which is used to identify a query as a common query or a novel query. In this work, Majority Vote Rule and Bayes Sum Rule are applied to implement the image query detection technique. For a common query, a relevant predefined image class will be predicted and within the class the relevant images are ranked. For hidden classes, during the retrieval process the features of the query image are extracted, then matched with the centroid of the each cluster. Among these clusters, features extracted from the query image that are nearest to the centroid of the cluster is selected. Then the query image is compared with the nearest images to the centroid of the selected cluster and the more relevant images are ranked.