A Survey of Semantic Gap Reduction Techniques in Image Retrieval Systems

摘要

An efficient and accurate image retrieval system is required to handle the increased usage of images. Low-level image features — color, shape and texture give the machine description of an image by describing its visual content which doesn’t match exactly with the high-level semantics of image. This mismatch corresponds to the semantic gap in image retrieval systems focusing only on low- level image features. They can satisfy the user’s demand of finding similar or relevant images only to some extent due to the semantic gap problem. Research focus needs to be shifted towards minimizing this semantic gap between the machine description and the human semantics of images. This paper presents a comprehensive review of the work done in reducing the semantic gap while retrieving images from the image database. The five state-of-the-art semantic gap reduction approaches: ontology, relevance feedback, machine learning, semantic template generation, and web based image retrieval are discussed in the paper. All the five techniques are evaluated on the basis of attributes- user involvement, offline/online processing, accuracy, iterative nature, time consumed, and search space reduction. These techniques can be used either individually or in hybrid manner to boost the performance and the relevance in image retrieval systems. The paper also highlights various ways of integrating these techniques together. 1. INTRODUCTION In view of the rapid progress and advancement of internet and digital technologies, a lot of multimedia data such as images, audio and videos, is used today as a part of our daily life. This data needs to be stored and retrieved in an efficient and effective manner. Many image retrieval systems are developed to serve this ongoing demand of retrieving images in image databases. They are used in many areas like medical, fashion, architectural designs, advertising, crime prevention, digital forensics, surveillance system and many more. The traditional approach for indexing, searching and retrieving images is based on manual text annotations, called Text Based Image Retrieval (TBIR) systems, where the images are first annotated manually by text or.