干旱区植物叶片识别研究

Study on Plant Leaf Identification in Arid Area

现有植物叶片识别方法都是基于扁平状叶片,而干旱区植物叶片多呈针叶,因此不适合干旱区植物叶片的识别,使得对于干旱区植物研究主要依靠专家识别,不利于对干旱区植物叶片的进一步研究。提出使用差异性值监督局部线性嵌入算法D-LLE,充分挖掘样本之间的类别信息,提高干旱区植物叶片的识别效率。首先利用金字塔梯度方向直方图(PHOG)的方法提取叶片图像特征,再使用PCA、LLE、WLLE、D-LLE等主流的降维算法,对提取的PHOG特征进行降维,最后建立支持向量机(SVM)的分类模型对植物叶片图像分类。经过这四种降维算法后的平均识别率分别为76.3%、85.3%、89.1%、95.5%;骆驼刺、苦豆子和沙枣的叶片正确识别率,相对其他植物叶片较低。通过实验证明了PHOG特征在植物叶片特征提取的可行性,使用D-LLE算法相比传统特征降维的算法具有更高的效率,且较适合于干旱区植物叶片的自动识别分类。

The existing plant leaf identification methods are based on flat- shaped leaves, while the arid areas of plant leaves are more needles, so it is not suitable for the identification of plant leaves in arid areas, so that the plant researches in arid area rely mainly on expert identification, not conducive to the further study of plant leaves in arid area. A Local Linear Embedding Algorithm D-LLE is proposed using the difference value to fully excavate the classification information between the samples to improve the identification efficiency of plant leaves in arid area. The Pyramid Histograms of edge Orientation Gradients(PHOG)is used to extract the leaf image features, then reduces the dimension of the extracted PHOG features by using the mainstream dimensionality reduction algorithm, such as PCA, LLE, WLLE, D-LLE. Finally, the classification model of Support Vector Machine(SVM)is established to classify the plant leaf images. The average recognition rates of these four dimensionality reduction algorithms are 76.3%, 85.3%, 89.1%, 95.5%. Respectively, compare with other plant leaves, the correct recognition rates of the leaves of camel prickles, kudou and zizyphus jujuba are low. This proves that PHOG is feasible in feature extraction of plant leaves by experiments. Furthermore, the D-LLE algorithm is more efficient than the traditional feature reduction algorithms and more suitable for automatic recognition and classification of plant leaves in arid area.