基于图像分割和决策层融合的未成熟柑橘检测方法  

Immature citrus detection using image segmentation and decision-level fusion strategy

作　　者：庄家俊 唐宇[1] 郭琪伟 何勇[2] 侯超钧 骆少明[1] 鲁斌 ZHUANG Jiajun;TANG Yu;GUO Qiwei;HE Yong;HOU Chaojun;LUO Shaoming;LU Bin(Academy of Contemporary Agricultural Engineering Innovations, Zhongkai University of Agriculture and Engineering, Guangzhou 510225, China;College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou 310058, China;Department of Intelligent Manufacturing, Wuyi University, Jiangmen 529000, China)

机构地区：[1]仲恺农业工程学院现代农业工程创新研究院,广东广州510225 [2]浙江大学生物系统工程与食品科学学院,浙江杭州310058 [3]五邑大学智能制造学部,广东江门529020

出　　处：《仲恺农业工程学院学报》2020年第4期46-52,共7页Journal of Zhongkai University of Agriculture and Engineering

基　　金：广东特支计划科技创新青年拔尖人才项目(2016TQ03N704);广东省科技计划项目(2016B020202008、2017A040405015和2017B010117012);广州市科技计划项目(201704020076、201904010206);江门市基础与理论科学研究类科技计划项目(2018JC01008)资助.

摘　　要：针对果园场景中未成熟柑橘色泽与背景目标相似度高、果实粘连的问题,提出一种改进GB色差图方法滤除输入RGB图像中的大量背景区域,采用数学形态学运算从保留的图像区域中重构前景和局部背景图像,并分别提取前景和局部背景区域标记,通过标记控制的分水岭变换进一步分离相互粘连/重叠的候选果实目标;提取局部二值模式纹理特征和方向梯度直方图形状特征描述候选目标;为提高检测结果的可靠性,结合支持向量机算法,通过双重特征决策层判别结果的逻辑与融合策略实现未成熟柑橘的检测.结果表明,该方法在验证集上的检测准确率和F1-measure指标分别达到了0.81和0.89,平均每帧图像误检率仅为0.02,能为未成熟柑橘果实的估产提供依据.The immature citrus fruits shared similar colour information with the other background participants under natural orchard environment and might encounter the phenomena of object adhesion.To this end,an improved GB chromatic mapping algorithm was presented to filter out the background regions from the input RGB images as many as possible.Based on the remaining regions,the foreground and local background regions were re-constructed using mathematical morphology operations,respectively,and the foreground and background markers were labelled from the re-constructed foreground and background images,respectively;the potential fruit regions of adhesion/overlap were further segmented using markers-controlled watershed transform;then,the texture and shape feature were extracted from those potential regions using local binary pattern(LBP)and histogram of oriented gradients(HOG),respectively.Finally,based on the extracted dual-modality features,a logic“and”operation fusion strategy by combining the decision results provided by support vector machines(SVM)was presented to locate the citrus fruits and further improve the reliability of the detection methodology.Experimental results on the validation dataset demonstrated that the detection accuracy and statistical indices F1-measure reached 0.81 and 0.89,respectively,with the average false detections of only 0.02 per image,indicating that the presented method could offer a reference for the yield estimation before citrus fruits'maturation stage.

关 键 词：未成熟柑橘 目标检测 色差图 分水岭变换 决策层融合 

分 类 号：S126[农业科学—农业基础科学] TP391[自动化与计算机技术—计算机应用技术]