基于机器视觉的中华稻蝗早期蝗蝻的识别和检测研究

发布时间：2018-01-25 06:23

本文关键词： 中华稻蝗早期害虫龄期识别机器视觉图像处理目标识别害虫检测 SVM DPM　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：水稻长期以来都是我国人民种植的主要粮食作物,当今人们对于粮食的安全和产量都提出双重要求的情况下,对稻田进行精细化管理,尽量减少化学药剂的使用是我们当今需要面对的问题。中华稻蝗占到稻田土蝗数量的80%左右,是稻田土蝗的优势种。目前对于中华稻蝗的预测和防治都还存在大量使用人力普查,盲目防治的问题。根据中华稻蝗在三龄前的群居特性,三龄后才扩散稻田的生态习性,本文着重探讨利用黏性陷阱诱捕中华稻蝗的前期蝗蝻,获取黏板图像,并在田间背景下采集中华稻蝗的图像,进行中华稻蝗的龄期识别、分类以及早期稻蝗的检测。(1)设计了一种针对中华稻蝗和稻田早期越冬害虫的图像获取系统。固定相机的高度和光圈以获得稳定的图像。实现了基于图像识别技术的中华稻蝗的龄期的识别和计数。将图像转化为HSV空间,利用V通道的灰度图像对原图的灰度图像进行增强,改善了不均匀光照、目标与背景灰度接近、目标较小的情况下Otsu法难以计算最佳阈值的缺陷。通过图像的增强和形态学处理对图像中的中华稻蝗进行分割,进而根据中华稻蝗不同龄期体态的阶跃性变化,完成对中华稻蝗的龄期识别和计数,为稻田蝗虫的虫情在线监测及管理决断提供依据。(2)提出了基于图像特征提取的早期中华稻蝗害虫和典型越冬害虫的识别方法。采用"图像分割——特征提取——分类器设计"的技术路线,在图像分割阶段,将早春时期的越冬害虫二化螟和稻纵卷叶螟与中华稻蝗的幼虫进行分割并将图像填充为相同大小;介绍了模式识别的主要方法,进一步针对图像分割后的害虫目标,提取了形态、不变矩、颜色和纹理等全局特征;基于图像的全局特征建立了 SVM分类模型,针对早春的越冬害虫和蝗蝻进行识别,识别准确率达到了 88.3%。(3)将DPM引入中华稻蝗的检测。类似于人体检测,害虫由于姿态变化多样,所处的环境复杂,对于害虫的检测也是一个很有挑战的课题。目前对于人体的检测已经有很多的方式,其中DPM(可变型部件模型)是在人体检测中效果较好的一种方法。本文研究了基于DPM的中华稻蝗检测。详细介绍了DPM的原理,将中华稻蝗的自然环境样本图像作为训练样本,应用训练的DPM模型对自然环境下获取的中华稻蝗图像以及黏板上的中华稻蝗图像进行检测。应用DPM算法,实现对大田复杂背景下稻蝗的检测。
[Abstract]:Rice has long been the main food crop planted by the people in our country. Nowadays, people have put forward double requirements for the security and yield of grain, and fine management of rice field has been carried out. Reducing the use of chemical chemicals as far as possible is a problem we need to face nowadays. Rice locust accounts for about 80% of the population of rice locust in paddy field. It is the dominant species of locust in rice field. At present, there are still a lot of problems in forecasting and controlling rice locust by manpower survey and blind control. According to the population characteristics of Chinese rice locust before the third instar. The ecological habit of rice field was diffused only after the third instar. This paper focused on the use of sticky trap to trap the nymph of the Chinese rice locust, to obtain the image of the slime plate, and to collect the image of the Chinese rice locust in the field background. The age of Chinese rice locust was identified. Classification and early Detection of Rice Locust. An image acquisition system for rice locusts and early winter pests in rice fields was designed. The camera height and aperture were fixed to obtain stable images. Count. Convert the image to HSV space. The grayscale image of the original image is enhanced by using the grayscale image of the V channel to improve the uneven illumination and the gray level of the target is close to the background. When the target is small, the Otsu method is difficult to calculate the defect of the optimal threshold. Image enhancement and morphological processing are used to segment the Chinese rice locust in the image. Then according to the step changes of the different ages of the Chinese rice locust, the identification and counting of the age of the Chinese rice locust were completed. To provide the basis for on-line monitoring and management decision of locust in rice field. Based on image feature extraction, a recognition method of early rice locust pests and typical overwintering pests was proposed, and the technical route of "image segmentation-feature extraction-classifier design" was adopted. In the stage of image segmentation, the overwintering insects, the Chilo suppressalis and the rice leaf roller, were segmented with the larvae of the Chinese rice locust, and the image was filled into the same size. The main methods of pattern recognition are introduced, and the global features such as shape, invariant moment, color and texture are extracted for the target of insect pests after image segmentation. Based on the global features of the image, a SVM classification model was established to identify winter pests and nymphs in early spring. DPM was introduced into the detection of Oryza chinensis. Similar to human detection, insect pests are in a complex environment because of the variety of attitude. Detection of pests is also a very challenging topic. There are many ways to detect human body. DPM (variable component Model) is a good method in human detection. This paper studies the detection of Chinese rice locust based on DPM. The principle of DPM is introduced in detail. The natural environment sample images of Chinese rice locust were used as training samples. The trained DPM model was used to detect the images of the Chinese rice locust in the natural environment and on the sticky plate, and the DPM algorithm was used to detect the rice locust in the complex background of the field.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;S435.112.4

【相似文献】