基于超像素特征的苹果采摘机器人果实分割方法  

Fruits Segmentation Method Based on Superpixel Features for Apple Harvesting Robot

作　　者：刘晓洋[1] 赵德安[1] 贾伟宽 阮承治 姬伟[1] LIU Xiaoyang;ZHAO Dean;JIA Weikuan;RUAN Chengzhi;JI Wei(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212013,China;School of Information Science and Engineering,Shandong Normal University,Ji’nan 250358,China;School of Mechanical and Electrical Engineering,Wuyi University,Wuyishan 354300,China)

机构地区：[1]江苏大学电气信息工程学院,镇江212013 [2]山东师范大学信息科学与工程学院,济南250358 [3]武夷学院机电工程学院,武夷山354300

出　　处：《农业机械学报》2019年第11期15-23,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(31571571、61903288);山东省自然科学基金项目(ZR2017BC013);福建省自然科学基金项目(2018J01471);江苏省高校优势学科建设项目(PAPD)

摘　　要：针对苹果采摘机器人在自然环境下对着色不均匀果实的识别分割问题,提出了基于超像素特征的苹果采摘机器人果实分割方法。首先,采用简单线性迭代聚类算法将图像分割成内部像素颜色较为一致的若干超像素单元;然后,提取每个超像素的纹理和颜色特征,并采用支持向量机将超像素分为果实和背景两个类别;最后,根据超像素之间的邻接关系对分类结果进行进一步修正。实验表明,该方法能够对大部分超像素单元进行正确分类,平均每幅图像被错误分类的超像素约为2.28个。与采用像素级特征的色差法和采用邻域像素特征的果实分割方法相比,采用超像素特征的果实分割方法具有更好的分割效果。在进行邻接关系修正前,该方法图像分割准确率达0.9214,召回率达0.8565,平均识别分割一幅图像耗时0.6087 s,基本满足实时性需求。In order to segment uneven colored apple fruits in natural environment,the fruit segmentation method based on image features extracted from superpixels was proposed for apple harvesting robot.Firstly,simple linear iterative clustering(SLIC),which was one of superpixel clustering algorithm was employed to segment original images into a set of superpixels.The color of pixels in the same superpixel was uniform relatively.Then,the color and texture features of superpixels were extracted.According to combined feature vectors,these superpixels were classified into fruit class and non-fruit class by support vector machine(SVM).Finally,the classification results were modified based on the adjacency relation of superpixels.The segmented fruits were made up of a set of superpixels in fruit class.The experiment results showed that the proposed method can classify a majority of superpixels and there were average of 2.28 superpixels in one image were classified falsely.Compared with the segmentation method based on pixel-level features and the segmentation method based on features of neighborhood pixels,the proposed method based on superpixel features had a better performance on fruit segmentation.The experiment of image segmentation with 100 images indicated that the precision and recall of proposed method can reach 0.9214 and 0.8565 respectively before modifying classification results.The running time of proposed method was 0.6087 s per image.

关 键 词：苹果 采摘机器人 特征提取 超像素分割 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]