Invariant scene description based on salient regions for preattentive similarity assessment

The majority of content based image retrieval systems use low-level features for similarity assessment of image content. However low-level descriptors are often not sufficient for encoding stably more complex image information. This paper proposes an intermediate level descriptor which captures the organization of the image scene based on salient regions of the image. The descriptors are invariant to specified image transformations. A non-linear combination scheme is exploit to combine the different descriptor models for preattentive similarity assessment where neither learning nor user-feedback is required. Results on a database containing more then 6,000 images are presented