Optimal Interactive Content-Based Image Retrieval

No doubt, the performance of a Content-Based Image Retrieval (CBIR) system depends on a) how efficient the image visual content is represented and b) the degree of importance, which is assigned to each content-descriptor. In the first case, efficient visual representation is achieved, apart from the extraction of appropriate descriptors, through a proper organization of them [1]. The second case faces the problem arising from the subjectivity of humans, which often perceive the same visual content in a different way. This is, for example, the case when a person is interested in color information of an image, while another one, or even the same under different circumstances, in texture or motion information. Furthermore, even if both people are interested in the same type of indexing, e.g., color indexing, they may interpret the image content quite different. For instance, the first person may want to search for a particular object with specific color characteristics, while the second for the global brightness of an image. To overcome the aforementioned difficulty, the human can be considered as a part of the retrieval process resulting in an interactive CBIR scheme [2]. In particular, the retrieval results are evaluated by the user and a degree of relevance is assigned to some selected images so that the system response is adapted according to the user's demands. This interactive framework is usually called relevance feedback, similarly to the definition used in traditional text-based information systems [3]. Towards this direction, some CBIR systems incorporating relevance feedback algorithms have been proposed. In particular, a relevance feedback algorithm based on a probabilistic framework was reported in [4], while a relevance feedback algorithm, which exploits the variation of the feature elements to perform the weight updating, has been investigated in [2] and [5]. Extension of [5] to negative examples has been presented in [6] very recently, while in [7] the extension is performed to the interaction of different feature elements with each other. However, as mentioned in the conclusions of [5], there is a need for an optimal weight updating strategy. In this case, optimality means that the system should be adapted so that its response is as close as possible to all relevant selected images and as far as possible to all irrelevant images. In this paper, an efficient optimal relevance feedback mechanism is presented to improve the performance of a content-based image retrieval system. In particular, the algorithm is based on an optimal weight updating strategy, which regulates the importance of each image descriptor to the retrieval response. The weights are updated so that the correlation between the query image and all selected images of high degree of relevance is maximized, while simultaneously the correlation over all selected images of low degree of relevance is minimized. The optimization results in an efficient scheme both in computational complexity and performance for the weight updating. Convergence of the weight updating is achieved if the user submits "consistent" relevant or irrelevant images to the system, i.e., images of similar content. This means that the system response adapts its performance to the user's information needs in a stable manner.