Abstract: Leaf margin is one of the main characteristics to identify plant species. Compared to leaf shape features, leaf margin features are much more subtle, so they are often indispensable in multiscale recognition of plant species as either dependent features or supplements for others. The progresses include designing 7 new margin feature descriptors, taking hierarchical classification organized by some semantic dictionaries to reach a better classification accuracy, and finally deciding plant species of a leaf node member by similarity evaluation and retrieval. Our experiments have revealed that the descriptors, named as the ratio of residual convex to leaf area and the ratio of right edge to left edge, are efficient to distinguish between different nonlobedleaf species and different nonintegrifoliousleaf species; the mean value of residual convex etc., is of other examples of useful descriptors to the identification between different nonintegrifoliousleaf species. By using the hierarchical classification in the feature space of multi leaf margin descriptors, 30 nonlobedleaf species have been divided into several leaf nodes, and the mean overall accuracy is better than 81.21%. The test of assessing the similarity between the new assigned leaf node member and the known samples has further demonstrated that the framework of jointly using the hierarchical classification and the image retrieval is effective for the identification of plant species.

Key words: leaf margin features, descriptors, residual error of convex, semantic dictionary, image retrieval