Semantic image retrieval is one of the key technologies to find useful multimedia information more efficiently on Internet or in multimedia database. Extraction of main regions in an image is a precondition for semantic image retrieval. In this article

an automatic approach to extract those main regions is proposed. It first partitions an image into fixed sized blocks

and an elementary segmentation is achieved by clustering the visual features of all the blocks of the image. Then the result of the original segmentation is improved by some extra processing. After that

a special method is employed to distinguish the foreground regions and background regions. Finally

the regions

which are considered not important to the image content

are eliminated

and it is done by analyzing the importance of every region. Our experiments for outdoor images containing relatively salient objects show that

the approach proposed in this paper can get rid of lots of information

which are not related to the image content

and at the same time can also reserve the main useful information for image semantics. It gives a better foundation for the further applications such as image retrieval and image understanding.