基于深度学习的重叠柑橘分割与形态复原  

Overlapping Citrus Segmentation and MorphologicalRestoration Based on Deep Learning

作　　者：黄磊磊 苗玉彬[1] Huang Leilei;Miao Yubin(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240

出　　处：《农机化研究》2023年第10期70-75,共6页Journal of Agricultural Mechanization Research

基　　金：上海市科技兴农攻关项目(2019-02-08-00-08-F01118);国家自然科学基金项目(51975361)。

摘　　要：自然场景下柑橘果实的遮挡和重叠现象是自动化采摘面临的一大难题,对此提出了一种基于深度学习的多阶段分割与形态复原方法。首先,通过引入Pointrend分支的Mask R-CNN完成对柑橘的识别及边缘细化的实例分割;其次,基于编解码器结构提出柑橘果实形态粗复原模型,并设计局部惩罚损失函数及交并比形状损失函数;再次,通过机器视觉方法,根据粗复原结果提取感兴趣区域;最后,通过基于部分卷积的形态精复原模型完成重叠柑橘的形态复原。实验结果表明:该方法相比传统方法具有更高的果实定位精度和更好的复原效果,能够为自然场景下的柑橘智能采摘提供参考。The occlusion and overlap of citrus in natural scene is a difficult problem for automatic picking.In this paper,a multi-stage segmentation and morphological restoration method based on deep learning was proposed for overlapping citrus in natural scene.Firstly,Mask R-CNN with Pointrend branch was applied for recognition and instance segmentation of citrus.Secondly,a rough morphological restoration model based on the encoder-decoder structure was proposed for citrus with the designing of local penalty loss and IoU shape loss.Then,the region of interest was extracted according to the rough restoration results using machine vision methods.Finally,the morphological restoration of overlapping citrus was completed by the morphological fine restoration model based on partial convolution.The experimental results show that the method presented in this paper has higher positioning accuracy and better restoration effect than traditional methods.And it can provide a reference for intelligent citrus picking in natural scene.

关 键 词：重叠柑橘 机器视觉 实例分割 形态重建 

分 类 号：S225.93[农业科学—农业机械化工程] TP391.41[农业科学—农业工程]