基于叶缘特征的植物图像分类检索

Plant image classification and retrieval based on leaf margin features

叶缘是植物属种识别分类可以参照的主要特征之一.与叶片形状特征相比,叶缘描述了尺度更细微的特征,对于弥补叶形识别特征的不足、以及从多尺度识别植物属种都有重要意义.在借鉴已有研究成果的基础上,设计了7个新的叶缘特征描述符、提出了以语义字典组织植物属种多层分类中的层间继承关系、以及通过叶节点成员相似性评估确定植物属种的技术框架和方法.通过分支结点描述符组合分类实验,证明了叶缘凸残差与叶局部面积比、右边长与左边长比对划分不同的非裂叶植物、以及划分不同的非全缘叶植物有效;叶缘凸残差均值等描述符对于划分不同的非全缘叶植物有效.通过多描述符组合的多层分类将30种非裂叶植物划分到多个叶节点,平均全局精度优于81.21%.而叶节点成员属种概率评估实验,进一步论证了这种多层分类和相似性检索框架的合理和有效性.

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 multiscale 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 spe-cies; 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.