基于Mask R-CNN的葡萄叶片实例分割  

Instance Segmentation of Grape Leaf Based on Mask R-CNN

作　　者：乔虹 冯全 赵兵 王书志[2] QIAO Hong;FENG Quan;ZHAO Bing;WANG Shu-zhi(School of mechanical and electrical Engineering,Gansu AgriculturalUniversity,Lanzhou Gansu 730070,China;Electrical Engineering College of Northwest University for Nationalities,Lanzhou Gansu 730070,China)

机构地区：[1]甘肃农业大学机电工程学院,甘肃兰州730070 [2]西北民族大学电气工程学院,甘肃兰州730070

出　　处：《林业机械与木工设备》2019年第10期15-22,共8页Forestry Machinery & Woodworking Equipment

基　　金：国家自然科学基金项目(61461005);甘肃省科技重大专项计划项目(1502NKDF023);甘肃农业大学研究生重点课程建设项目(GSAU-ZDKC-1804)

摘　　要：在大田环境下对葡萄生长状态和病虫害进行长期动态自动监测,需要对监控摄像头拍摄的每张单个叶片进行实例分割,工作量大,为解决这一问题采用了基于MaskR-CNN的实例分割算法。该算法是在Faster R-CNN的基础上增加一个能在候选区域上进行分割任务的分支,葡萄叶片图像首先通过区域卷积神经网络生成候选区域,利用FastR-CNN的卷积层提取葡萄叶片的整体特征,得到特征图;再由ROIAlign对特征图进行像素校正,并对每一个ROI预测,得到其类别及预测框,每一个ROI再通过一个全卷积网络对每个像素进行分类和分割。对不同天气下正常的葡萄叶片、病害叶片以及不同品种的葡萄叶片图像进行分割试验,结果表明,本算法对正常叶片、病害叶片及不同品种叶片分割的平均精度(average precision,AP)分别是0.9108、0.9068、0.9044、0.8845、0.9028。该方法对不同天气及复杂背景下的叶片实例分割都具有较好的鲁棒性和较高的精度。Aiming at the problem of long-term dynamic automatic monitoring on grape growth state and diseases in the field environment,it is necessary to do instance segmentation for each blade of the image taken by surveillance cameras.This paper solved this problem by using the instance segmentation algorithm based on Mask R-CNN.This algorithm added a branch which can perform segmentation task on the candidate region on the basis of Faster R-CNN.First,the grape leaf image generated the candidate region through the regional convolution neural network,then extracted the whole characteristics of the grape leaf by using convolution of Faster R-CNN and feature map is obtained.Then,RoIAlign performed pixel correction on the feature map and predicted each ROI,getting its category and prediction box.Each ROI classifies and segments each pixel through a fully convolution network.In this paper,the segmentation test of healthy grape leaf images,diseased leaf images and grape leaf images of different varieties in different weather conditions is carried out.The experimental results showed that the average precision(AP)of the algorithm for normal leaves,diseased leaves and different varieties leaves is 0.9108,0.9068,0.9044,0.8845,0.9028.This method has better robustness and higher precision for instance segmentation on grape leaf under different weather and complex background.

关 键 词：MaskR-CNN 实例分割 复杂背景 天气条件 葡萄叶片 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]