基于多光谱图像和SVM的玉米田间杂草识别

Corn Field Weeds Recognition Based on Multi-spectral images and SVM

为解决变量喷洒对杂草识别速度与正确率的要求,提出了一种基于多光谱图像和SVM的杂草识别新方法。通过多光谱成像仪获得玉米与杂草图像,采用IR-R的多光谱融合并结合Otsu分割法完成背景分割;随后对植被图像进行目标分割与形态学处理,提取出所有植被叶片图像,在此基础上提取了叶片11个形状特征参数和纹理特征参数。为提高算法的实时性,对叶片的特征参数进行主成分分析,将前3个主成分作为支持向量机的输入建立模式识别模型。结果表明,降维后对于未知预测样本的识别正确率达到85%,用时0.001 415s。与直接利用支持向量机的90%的识别率和0.105 165s的用时相比,该算法在满足识别率的同时,用时更少,为田间杂草的快速识别提供了一种新方法。

To solve the requirement of variable spray for weed identification speed and accuracy rate,proposed a new method based on multi-spectral images and SVM weed identification.By the corn and weed images of multi-spectral image,IR-R multi-spectral fusion and combination of Otsu segmentation method was used to complete the background segmentation.Then vegetation image object segmentation and morphological processing was taken before extract all the vegetation leaf images.Based on this,11 leaf characteristic parameters of shape and texture was extracted.To improve the real-time,principal component analysis was taken for the characteristic parameters of leaves,and the first-three principal components was taken as input of support vector machines to establish pattern recognition model.The results showed that 85% recognition accuracy for unknown prediction samples after dimensionality reduction,with a time of 0.001 415s,compared with recognition rate of 90% and 0.105 165s of the direct use the support vector machine.The algorithms cost a little more time to meet the require recognition accuracy,and provides a new method for the rapid identification of weed.