A New Approach to Object-Related Image Retrieval

In Journal of Visual Languages and Computing Vol.11, No. 3, pp. 345-363, 2000

Abstract:

Image retrieval has been commonly attempted using non-semantic approaches. It is clear though, that semantic retrieval is more desirable because it facilitates the user's task. In this paper, we present a new approach to semantic access of a database of images by asking for the presence of certain objects; this is known as object-related image retrieval. This approach is built within a classical computer vision framework (i.e. localization, segmentation and identification). Our approach first searches for the main areas of attention (most salient areas of an image) and then apllies appearance-based methods to classify (index) all images by ``symbolic" names. These names are refered to objects, which finally allows the use of semantics driven by these object names, e.g. retrieve ``all those images that have a bull and Melissa's face". The use of a totally automatic system would cause some errors of indexing (and so retrieval). To solve this we use a human-in-the-loop strategy where a human expert is placed after the two outputs of the system to confirm their ``correctness". An experimental result using a database of $3000$ images is presented.