基于SURF算法的大豆灰斑病视觉识别系统

Visual Identification System of Soybean Frogeye Leaf Spot Based on SURF Feature Extraction

机器视觉技术是农田信息采集系统的关键技术之一,在精细农业中有广泛应用,农作物病害部位的精准识别作为精准施药的前提和关键,其识别精准度对病害防治效果影响巨大,当前农田信息采集系统识别精度差的问题急需解决。本研究结合机器视觉与计算机图像处理技术,以Visual Studio 2010为开发平台,搭建大豆灰斑病计算机视觉识别系统。由计算机摄像头采集大豆叶片图像,通过对彩色图像灰度化预处理,分别对比Open CV(Open Source Com-puter Vision Library)计算机视觉开源库中两个图像特征检测识别方法—SURF(Speeded Up Robust Features)法和SIFT(Scale-Invariant Feature Transform)法,对图像灰斑特征点进行检测,两种算法在输出帧率上差别明显,SIFT算法输出帧率为0. 3～0. 5 fps,SURF算法输出帧率为0. 6～0. 9 fps,考虑设备性能和灰斑识别的精准性,最终选用SURF算法;建立图像像素点海森矩阵,经高斯滤波,利用非极大值抑制法确定特征点,再由立体空间差值法定位极值点,根据哈尔小波响应值选取特征点主方向,构造SURF特征点描述算子以提取特征点,编写相关程序代码,分别对黑农44分枝期和结荚期大豆叶片灰斑进行检测。结果显示:分枝期叶片纹理少,特征点少,检测效果好,检测正确率97. 28%,耗时0. 97 s,结荚期叶片纹理增多,绘制特征点较多,检测正确率89. 49%,耗时1. 19 s,基本满足大豆灰斑病识别系统功能需求。通过田间试验,利用FLANN算法对分枝期视频图像进行特征点提取,实现对视频图像帧检测并匹配特征点的目的,检测率为90. 7%,匹配率93. 8%,该大豆灰斑病视觉识别系统的构建能够为下一步精准施药及相关农田信息精准采集系统设计提供思路与参考。

Machine vision technology is one of the key technologies of farmland information collection system,which is widelyused in precision agriculture. As premises and keys of precision pesticide application,recognition accuracy of crop disease sitehas strong influences on the diseases control. The badness of recognition accuracy is urgent to be solved. Combining machinevision and computer image technology,visual studio 2010 is used as the experimental platform to build a computer vision iden-tification system of soybean gray spot. The image of soybean leaf was collected by computer camera,and the color image waspreprocessed by graying. Two image feature detection methods from Open CV(Open Source Computer Vision Library)-SURF(Speeded Up Robust Features) algorithm and SIFT(Scale-Invariant Feature Transform) algorithm were used to detect grayspot feature points. There were obvious differences between two methods in output frame rate: The range of SIFT algorithmoutput frame rate was between 0. 3 and 0. 5 fps,and SURF algorithm was between 0. 6 and 0. 9 fps. Considering the efficiencyand equipment performance,SURF algorithm was selected. The Hessian Matrix of pixel points in the image was established.The feature points were determined by using the Gauss filter and the non-maximum suppression,and then the extreme pointswere located by the spatial interpolation. The main direction of the feature points was selected according to the Hal wavelet.The descriptor of SURF feature point was constructed to extract feature points,and the program codes were compiled to detectthe gray spot in soybean leaves of Heinong 44 in branching stage and pod bearing stages respectively. The results showed thatthe leaf texture and the feature points in branching stage were few,detection effect was good,the correct rate of detection was97. 3%,and detected time was 0. 97 s. The leaf texture and the feature points increase in pod bearing stages,the correct rateof detection was 89. 49% and detected time was 1. 19 s,which basically meets the functional requirements of soybean grayspot recognition system. Through field trials and using FLANN algorithm to extract feature points of video images,detecting ofvideo image frame rate and matching of feature points were achieved. The detection rate was 90. 7% and the matching rate was93. 8%,which provides ideas and references for the next research on precision pesticide application and the design of otherfarmland information collection system.