【作者】 肖健；

【导师】 曾爱军； 何雄奎；

【作者基本信息】 中国农业大学 ， 农业机械化工程， 2005， 硕士

【摘要】 针对果树病虫害防治工作中劳动强度大、防治器械落后、农药污染严重等问题,提出将机器视觉技术应用于果园对靶喷雾系统。通过对仿真果树图像实时采集、分割获取树形和枝叶分布信息,并以十六进制编码向下位喷雾装置下达喷雾指导信息。对于每一帧图像(768×576象素)从采集到下达喷雾信息的总处理时间为0.579秒。 本文重点研究果树图像分割方法。首先,采取基于颜色特征的分割。在筛选颜色特征时,突破前人定性评价颜色特征优劣的方法,引入以“误判率”为指标的定量颜色特征评价方法。通过“误判率”曲线可以明确的说明各种颜色特征在分割中的表现,最终选定ExG1作为颜色特征。其次,以复杂环境中果树图像分割为目的,初步探讨了基于纹理特征的图像分割方法。研究中将图像分成无重复的纹理区块(16×16象素)。由区块纹理特征:能量、熵辅以图像区块X和Y坐标组成纹理区块样本向量代表该小区特征。对由土壤、果树枝叶、草地和天空四种果园中常见纹理合成的实验图像聚类,分类结果正确率达100%。未加象素区块位置信息,仅以能量和熵作为向量特征时,这四种纹理合成图像分类正确率从4%到90%不等。表明在纹理分析中,纹理本身特征辅以纹理区块位置信息可以提高分割效果。更多还原

【Abstract】 The heavy reliance on chemicals raise many enviromental and economic concerns, the triditonal sprayer and pesiticide application method could not suit the mordern orchard spraying. In order to reduce chemical use and get farmer free from hard work in orchard spraying, a simulated real-time machine vision based robotic orchard spraying system was builded, that include three parts: (1) real-time tree image catching system; (2) tree image preprocessing and recognition system; (3) image expression and message communication system. It took 0.579s for the hole system to processing one image (768 × 576 pixels) , representing a 3m by 3m region of one tree.The emphasis of this dissertation was laid on the study of tree image recognition. Firstly, color was used to distinguish tree crow from the background. For sake of finding the proper feature in color model, ultimate measurement accuracy(UMA) was introduced to decide which feature was appropriate. According this evaluation method, ExG1 was selected as the color feature in tree image segmentation. Secondly, textural feature analysis was used to identify tree crown. The image (768×576 pixels) was divided into rectangles(16×16 pixels). Energy and entropy deriving from GLCM and coordinate(x,y) of each rectangle were selected as index of vector which represent every rectangles. The man-made image which include four different texture (soil, leafs of fruit tree, grassplot and sky) are tested, the result showed that each kind of texture rectangel was recognised with 100%. But if the vector of rectangel without coordinate, the accurateness of recognition was different (from 4% to 90%). So the coordinate is very important to textural analysis.更多还原