基于机器视觉的玉米植株数量获取方法与试验

摘要

获得田间的玉米植株数量对于优化不同玉米品种的种植密度有重要意义，玉米植株数量也是计算新玉米品种平均每株产量的重要参数。为了减轻人工获得玉米植株数量的劳动强度，提高数据的准确率，该文利用基于机器视觉的图像处理技术来获得玉米植株数量。被留高茬玉米收获机作业之后的地块，有一定高度的玉米秸秆站立在地表，摄录这样的图像信息可以大大简化图像处理的难度，提高结果的精确度，所以将图像采集装置安装在留高茬玉米收获机之后来获得视频流。后处理过程中，将视频文件分解为图片文件，然后将真彩色的RGB图片文件转化成灰度图像进行图片的配准，再将灰度图像转化为二值图像进行图像分割与边界提取，最后找到玉米秸秆断面的几何中心并进行标记，统计标记结果即获得玉米植株数量。试验结果显示，人工播种与机械播种在图像识别的误差上没有显著差异（P＞0.05）；机器视觉识别出来的玉米植株数量与实际数量也没有显著差异（P＞0.05），其平均误差为6.7%；并且该误差不会随着图像中玉米植株数量的增加而产生积累。该文的设计可以降低机器视觉在识别玉米植株数量过程中的难度，提高图像识别的准确度，更好地服务生产实际问题。%It is very important to count corn population for optimizing plant density of each corn variety, and corn population is also a very important parameter for calculating average yield of each corn plant. Generally speaking, there are three methods to count corn population, which are based on mechanism, photoelectric technology and machine vision separately. In order to decrease the labor intensity and improve the accuracy, image identifying technology is used in this paper to obtain corn population. As corn seedling and weeds have some similarities, and not every corn seedling can grow up to a ripe corn, counting ripe corn’s population is more significant than counting corn seedling’s population. But it is not easy to enter the ripe-corn field for machinery, additionally, corn leaves will overlap and be blown by even slight wind, which will disturb image obtaining. There are also some solutions for the problems mentioned above, for example, corn fields will have a big difference after being operated by high-stubble corn harvesters. A section (300 to 500 mm) of corn stalks will be retained in the field after being harvested by high-stubble corn harvester, and there will be a distinct comparison between the stubble cross-section’ color and other sceneries in the image. Processing images obtained from these fields will decrease the difficulties and improve the accuracy. So image acquisition equipment is mounted on the high-stubble corn harvester. Actually, the visual document obtained from the field is video document at first, and then the video document is decompressed into image. Subsequently, the RGB (red, green, blue) images are converted to gray images for mosaicking. The gray images are converted to binary images in the image segmentation and border extraction section next to image mosaicking section. Although the cross-section of stubble is not a perfect circle, its edge has an obvious feature compared to other objects in the image. At last, a function is used to extract the edge of stubble cross-section, and then the centroid of cross-section is marked. So corn population can be obtained by counting the marks. Experiments were done to test the method and the design in this paper was in autumn of 2013. Experiment results have expressed that there is no significant difference (P<0.05) between artificial seeding and mechanical seeding; and there is also no significant difference (P<0.05) between automated counting and manual counting. The automated count’s mean error is only 6.7%, and this error will not accumulate along with the increasing number of corn plant. The results of artificial count and automated count are linear correlation. The results of linear regression analysis show that the values ofR2 of four experiments are 0.95, 0.90, 0.91 and 0.91, respectively, the slopes of four regression lines are 0.93, 0.91, 1.08 and 0.95 separately, and the intercept of four regression lines are 0.98, 0.97, -0.12 and 0.97 respectively. The design in this paper can reduce the difficulty in identifying corn stalks in images, and improve the image-identifying accuracy at the same time, and hence can better serve the real problems in counting corn population.