The research field of “Content-Based Image Retrieval (CBIR)” is closely related to several others. This chapter provides an overview of the most relevant research fields and their interrelationship regarding this investigation. For each one, a summary of recent, related research is presented. In addition, the related preliminary work of the author is shortly presented. Based on this background information, major challenges in CBIR are discussed. The scope and the aims of this investigation have been adjusted to accommodate those challenges with respect to the given resources.
TopMany of the basic issues in CBIR have been collected by Eakins and Graham (Eakins & Graham, 1999) in 1999. The most severe problems of image retrieval identified in their report remain unsolved. The key to image retrieval is “bridging the semantic gap” between low-level image content (pixels, as seen by machines) and its high-level meaning (semantics, as seen by humans) (Zhao & Grosky, 2002). In Figure 1, the semantic gap is represented by the two research fields Features/Similarity and Annotation. The first one deals with low-level content whereas the second one deals with its high-level counterpart. A technology to connect these two fields is the Categorization.
Figure 1. Research field dependencies
New technologies and methods have been continually improving the quality of MIR (Wang et al., 2007) and several real-life applications have been developed in many areas. Yet, a large amount of recent challenges in MIR are still to be solved. Lew et al. (Lew et al., 2006). provide a summary of these challenges.
Currently the research seems to shift from image retrieval towards video retrieval. Nevertheless it should be considered that the diversity in multimedia retrieval is beneficial for all sub-disciplines as they overlap in many cases. Also the researchers coming from several different disciplines contribute to the richness of different approaches and solutions (Wang et al., 2006).
The recent overview provided by Datta et al. (Datta, Joshi, Li, & Wang, 2008) concludes with the statement that the research focused more in systems, feature extraction and relevance feedback than in application-oriented aspects such as interface, visualisation, scalability and evaluation. Thus it seems desirable to improve research in these areas.
According to Lew et al. (2006) there are several recent research topics trying to bridge the semanticgap. In human-centred computing the system tries to satisfy the user while keeping the interface easily understandable. Learning algorithms could be beneficial in adding semantic value. Developing new features based on low-level information may still be beneficial, when adapted to the human perception and easy to use. Their conclusion is that none of the major challenges in MIR are actually solved and all areas require significant further research.
This section briefly describes several research areas closely related to CBIR (Figure 1). Some of these are user-centred and some deal with the underlying methodology. Concerning the user interface, the areas of Browsing, Query Languages and Relevance Feedback play an important role. This supports the user in creating queries and to navigate through a given repository. Creating Feature Vectors and Similarity Measures is a matter of improving the system quality on the server side. Similarly, the Annotation is important to enable keyword based retrieval. They usually need to be tuned by an Evaluation process. The area of Categorization is an approach to link existing low-level features directly to high-level keywords and categories. Finally, there are Retrieval Framework Designs available, which provide researchers with the basic functionality needed to do this kind of research.