ACCURATE IMAGE RETRIEVAL BASED ON COMPACT COMPOSITE DESCRIPTORS AND RELEVANCE FEEDBACK INFORMATION

In this paper a new set of descriptors appropriate for image indexing and retrieval is proposed. The proposed descriptors address the tremendously increased need for efficient content-based image retrieval (CBIR) in many application areas such as the Internet, biomedicine, commerce and education. These applications commonly store image information in large image databases where the image information cannot be accessed or used unless the database is organized to allow efficient storage, browsing and retrieval. To be applicable in the design of large image databases, the proposed descriptors are compact, with the smallest requiring only 23 bytes per image. The proposed descriptors' structure combines color and texture information which are extracted using fuzzy approaches. To evaluate the performance of the proposed descriptors, the objective Average Normalized Modified Retrieval Rank (ANMRR) is used. Experiments conducted on five benchmarking image databases demonstrate the effectiveness of the proposed descriptors in outperforming other state-of-the-art descriptors. Also, a Auto Relevance Feedback (ARF) technique is introduced which is based on the proposed descriptors. This technique readjusts the initial retrieval results based on user preferences improving the retrieval score significantly. An online demo of the image retrieval system img(Anaktisi) that implements the proposed descriptors can be found at .