Abstract: For the purpose of automatic surveillance on wine grape in some aspects such as growth state, disease and insect pests, grape leaves should firstly be detected in an image. In this paper, a method is proposed in which Histogram of Oriented Gradient（HOG） features of a leaf are extracted and a classifier trained by support vector machine is used to identify the grape leaves in the images. The locations of the leaves in the image are determined by multi-scale object localization and mean shift algorithm. The experiment results show that the recognition rate of grape leaves and four common weed leaves reaches 95.5% when using the classifier trained with linear kernel function. The proposed method is robust to variety of illumination and environment. The detection rate of grape leaves of the images under natural conditions is over 80% in experiment.

Key words: Histogram of Oriented Gradient（HOG）, support vector machine, recognition, mean shift algorithm, location, detection

摘要： 在酿酒葡萄生长状态与病虫害自动监测中，需要在图像中检测出葡萄叶片，通过提取葡萄叶片图像的方向梯度直方图（HOG）特征投入到支持向量机（SVM）分类器中以实现对葡萄叶片的识别;结合多尺度目标定位和均值漂移算法还可以自动确定图像中葡萄叶片的位置。实验结果表明，使用线性核函数训练后的分类器对葡萄叶片和四种常见杂草的识别率达95.5%。该方法对光照和环境变化有较好的鲁棒性，自然条件下成像的叶片图像的葡萄叶片检出率达到了80%以上。

关键词: 方向梯度直方图（HOG）特征, 支持向量机, 识别, 均值漂移算法, 定位, 检测