基于稀疏实例与位置感知卷积的植物叶片实时分割方法
Real-time segmentation of plant leaves based on sparse instances and posi-tion aware convolution
任守纲 1朱勇杰 2顾兴健 2武鹏飞 3徐焕良2
作者信息
- 1. 南京农业大学人工智能学院,江苏 南京 210095;国家信息农业工程技术中心,江苏 南京 210095
- 2. 南京农业大学人工智能学院,江苏 南京 210095
- 3. 新疆兴农网信息中心/新疆维吾尔自治区农业气象台,新疆 乌鲁木齐 830002
- 折叠
摘要
植物叶片分割在高通量植物表型数据获取任务中起着关键作用.目前,多数植物叶片分割方法专注于提高模型分割精度,却忽视模型复杂度和推理速度.针对该问题,本研究提出一种基于稀疏实例激活与有效位置感知卷积的实例分割模型(ePaCC-SparseInst),实现植物叶片实时、精确分割.在ePaCC-SparseInst 中引入 1 组稀疏实例激活图作为叶片对象表示方式,并使用二部图匹配算法实现预测对象与实例激活图的一一映射,从而避免了繁琐的非极大值抑制(Non-maximum suppression,NMS)运算,提高了模型的推理速度.此外,在实例分支中引入有效位置感知卷积(ePaCC)模块,在增大模型全局感受野的同时提高了模型的推理速度.在Komatsuna数据集上,ePaCC-SparseInst平均分割精度(AP)达到 85.33%,每秒传输帧数达到 43.52.在相同训练条件下,ePaCC-SparseInst的性能优于SparseInst、Mask R-CNN、CondInst等实例分割算法.此外在CVPPP A5 数据集上,ePaCC-SparseInst较上述算法同样取得了更好的分割精度和推理速度.本研究提出的方法采用纯卷积的架构实现了叶片的实时分割,可以为在移动端或边缘设备上获取植物表型数据提供技术支持.
Abstract
The segmentation of plant leaves plays a crucial role in high-throughput plant phenotyping data acquisition tasks.Currently,most methods for plant leaf segmentation focus on improving the accuracy of the segmentation model but o-verlook the model's complexity and inference speed.In response to this issue,this study proposed an instance segmentation model(ePaCC-SparseInst)based on sparse instance activation and efficient position-aware convolution to achieve real-time and accurate segmentation of plant leaves.In ePaCC-SparseInst,a set of sparse instance activation maps was introduced as the representation of leaf objects.A bipartite graph matching algorithm was employed to establish a one-to-one mapping be-tween predicted objects and instance activation maps,thereby avoiding the cumbersome non-maximum suppression(NMS)operation and improving the model's inference speed.Additionally,an effective position-aware circulate convolution(ePaCC)module was introduced into the instance branch,which increased the model's global receptive field and enhanced its inference speed.On the Komatsu-na dataset,ePaCC-SparseInst achieved an average segmentation precision(AP)of 85.33%and an inference speed of 43.52 frames per second(FPS).Under the same training conditions,its performance surpassed instance segmentation al-gorithms such as SparseInst,Mask R-CNN,and CondInst.Furthermore,on the CVPPP A5 dataset,ePaCC-SparseInst a-chieved better segmentation accuracy and inference speed than the aforementioned algorithms.The proposed method used a pure convolutional architecture to achieve real-time leaf segmentation,which could provide technical support for obtaining plant phenotypic data on mobile or edge devices.
关键词
实例分割/计算机视觉/植物表型/叶片分割Key words
instance segmentation/computer vision/plant phenotypes/leaf segmentation引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据