An Effective Content-Based Image Retrieval System by Hierachical Segmentation

Abstract

As the inaccuracy of current image segmentation methods, it’s unavoidable for the objects with discrepant components to be segmented into different regions. As a result, good image retrieval performance could not be achieved by those region-based image retrieval approaches. Furthermore, the complexity of image segmentations is also an unmanageable issue in the scenarios with complex backgrounds. Aspired by the wavelet multi-resolution analysis, the objects with different scales, orientations, and locations, can be retrieved by their invariant features and hierarchical multi-resolution segmentations. For simplification, the hierarchical segmentation is conducted to segment one image into equal block with different shifts and sizes in one hierarchical way, and those blocks can form a complete pitch to the image at different hierarchical levels with different shifts and sizes. The smaller the sizes of blocks are, the higher the hierarchical levels. Then, the similar metrics of these sub-blocks to query image, are evaluated to retrieve those sub-blocks with contents in query images. Meanwhile, the location and scale information about query objects can also be returned in retrieved images. With geometric invariants, normalized histograms and their combinations as invariant features, the hierarchical segmentation-based image retrieval scheme are tested by experiments via one image database with 500 images. The retrieval accuracy with geometric invariants as invariant features can achieve 78% for the optimal similar metric threshold, only inferior to that of region-based image retrieval schemes, whose retrieval accuracy in our experiments is 80% with expectation maximized segmentations.

Work supported by Gansu Natural Science Foundation of China under Grant No. 3ZS051-A25-047.