Interactive Visual and Semantic Image Retrieval

One direct consequence of recent advances in digital visual data generation and the direct availability of this information through the World-Wide Web, is a urgent demand for efficient image retrieval systems. The disclosure of the content of these millions of photos available on the internet is of great importance. The objective of image retrieval is to allow users to efficiently browse through this abundance of images. Due to the non-expert nature of the majority of the internet users, such systems should be user friendly, and therefore avoid complex user interfaces. Traditionally, two sources of information are exploited in the description of images on the web. The first approach, called text-based image retrieval, describes images by a set of labels or keywords [1]. These labels can be automatically extracted from for example the image name (e.g. ’car.jpg’ would provide information about the presence of a car in the image), or alternatively from the webpage text surrounding the image. Another, more expensive way would be to manually label images with a set of keywords. Shortcomings of the text-based approach to image retrieval are obvious: many objects in the scene will not be labeled, words suffer from the confusions in case of synonyms or homonyms, and words often fall short in describing the esthetics, composition and color scheme of a scene. However, until recently many image retrieval systems, such as e.g. Google-image search, were exclusively text based. A second approach to image description is called content-based image retrieval (CBIR). Here users are provided with feedback from an image-query purely based