摘要
【目的】探索大范围、实时和准确监测柑橘园实蝇发生的技术。【方法】设计了基于机器视觉技术的橘小实蝇、南瓜实蝇和瓜实蝇识别算法。该算法包括:(1)基于颜色特征的目标区域分割,(2)基于Hough变换的盾形区域长轴搜索和图像配准,(3)基于BP神经网络的实蝇目标识别等主要部分。【结果】所有程序基于Matlab软件编写,试验中对自制样本库的120头橘小实蝇、南瓜实蝇和瓜实蝇图像样本进行配准,准确率为100%,平均耗时0.4 s;应用建造的BP神经网络模型识别3种实蝇的准确率为100%,取得了理想的识别效果。【结论】由于目标区域分割中阈值的选取受实验环境因素的影响,进一步研究的方向为提高阈值选取的自适应性。
【Objective】 For monitoring fruit flies in citrus orchard on large-scale, real-time and accurate.【Method】 The algorithm was developed to identify Bactrocera dorsalis, B. cucurbitae and B. tau(Dipetra:Tephritidae) based on machine vision technology. The fruit fly sample image library was produced and the body characteristics for distinguishing the three different species of fruit fly were analyzed. Three major steps were included in the identification algorithm as:(1) the image's target and background region segmentation were conducted under the YCbCr color space,(2) the fly's scutellum area's long axis searching and image registration were achieved using Hough transform, and(3) the BP neural network fruit fly identification model was established to distinguish the fruit flies in the images. Every algorithm was programmed on the Matlab software. 【Result】 The image registration algorithm was applied to 120 images each of which contained a single insect of B. dorsalis, B. cucurbitae or B. tau(Dipetra: Tephritidae) from the self-captured sample library. The BP neural network model was tested to identify the fruit flies. Experimental results indicated that:(1) the yellow scutellum at the waist and abdomen of three different kinds of fruit fly contained the largest yellow area within the body, thus it could be applied as the long axis searching area during image registration. The vertical yellow lines at the middle of the back chests of both B. cucurbitae and B. tau(Dipetra: Tephritidae) could be applied to distinguish these two kinds of fruit fly from the B. dorsalis. Further, the area rate of these lines to the whole body could be used to distinguish these three kinds of fruit fly.(2) The image registration accuracy was 100% and the average registration time consumption was 0.4 s.(3) The BP neural network model accuracy was 100%, indicating a satisfactory identification. 【Conclusion】 Since the threshold selection for target region segmentation was influenced by the experimental environment, further research will focus on improving the self-adaptive of the threshold selection.
出处
《果树学报》
CAS
CSCD
北大核心
2014年第4期679-683,751,共6页
Journal of Fruit Science
基金
国家现代农业(柑橘)产业技术体系建设专项(CARS-27)
广东省高等院校学科与专业建设专项(2013KJCX0032)
广州市科技计划(2013J2200069)
关键词
柑橘园
机器视觉
实蝇
HOUGH变换
精细农业
Citrus orchard
Machine vision
Fruit fly
Hough transform
Precision agriculture
