Identifying leaf shapes is a challenging task due to the small inter-class differences and large intra-class variations appeared in the leaf shapes.The self-occlusion and noise disturbing also make this task become very difficult.In this paper,a novel shape descriptor,termed Chordfeatures matrices (CFM),is proposed for plant leaf image classification and retrieval.In the proposed CFM,the chord of the object contour is partitioned into two parts,the inner one and the outer one,according to the region enclosed by the contour.Due to their high correlations to the convex and concave properties of the contour,their lengths are utilized to generate two matrices for characterizing the convex and concave properties of the contour in multiscale.In additional,the projection lengths from the arc to the chord are also used to build a matrix for reflecting the bending degree of the contour.The combination of these three matrices characterizes the contour from coarse to fine which results a shape descriptor with powerful discriminative ability for accurate plant leaf classification and retrieval.The proposed CFM has been tested on three Challenging plant leaf image databases,including the well-known the Swedish,Flavia and ImageCLEF databases.All the experimental results demonstrate that the proposed CFM approach outperforms the state-of-the-art methods for shape based leaf recognition.An extra experiment conducted on the MPEG-7 dataset further indicates its potential application to the task of general shape recognition.%植物叶片形状一般具有小的类间差异和大的类内变化,再加之叶片的自遮挡和噪声的干扰,给叶片形状的识别带来了很大的挑战.文中提出了一种新的形状描述子——弦特征矩阵(Chord-Features Matrices,CFM),精确而又鲁棒地应用于植物叶片图像的分类和检索问题.该方法将目标轮廓线的弦,依据轮廓线所围成的区域,分成内和外两个部分.因其与轮廓线凸凹特性的相关性,该方法用多个尺度级的内部弦长和外部弦长生成两个矩阵,旨在隐式地描述轮廓线的多尺度的凸凹特性.该方法还定义了多个尺度级的弧到弦的平均投影长度,并构成矩阵,以反应轮廓线在各个尺度级下的弯曲程度.组合这3个矩阵所形成的CFM描述子,全面地刻画了轮廓线的几何特性,具有强的形状表征能力.用Swedish、Flavia和ImageCLEF这3个具有挑战性的叶片图像测试集,对文中提出的CFM方法分别进行分类和检索性能评估.实验结果表明,文中提出的方法在精确度和对噪声的鲁棒性方面均优于其他植物叶片图像分类和检索方法.而文中提出的方法在MPEG-7测试集上的实验结果则验证了其具有应用于一般的形状识别任务的潜力.
展开▼
