高级检索+

基于多视角时间序列图像的植物叶片分割与特征提取

Segmentation of Plant Leaves and Features Extraction Based on Muti-view and Time-series Image

  • 摘要: 为了解决多种类植物在生长过程中不同时间点动态变化表型参数提取困难问题,提出了一种基于多视角时间序列图像和深度卷积神经网络Mask-RCNN的植物茎叶实例分割方法,在拟南芥、玉米和酸浆属3种代表性植物上进行了实验。结果表明,训练得到的基于Mask-RCNN的植物分割模型对在不同生长时期的植物茎叶的识别精度(mAP0.5)大部分在70.0%以上,最高可以达到87.5%,模型通用性较好。同时,针对茎叶遮挡问题提出的基于多视角图像的跟踪算法,可进一步提高植物茎叶参数提取的准确率。本文提出的以茎叶为代表的植物器官分割和特征提取方法具有性能高效、成本低、通用性和扩展性好的优势,可为不同场景下植物全生长过程中的多表型参数提取提供参考。

     

    Abstract: Phenotyping aims to measure traits of interest and a key part of this requires the accurate identification of defined parts of the organism. Instance segmentation of organs, such as leaves, is a crucial prerequisite for plant phenotyping. Firstly, whether deep learning methods(such as Mask-RCNN) had generality for leaf and stem segmentation was evaluated. Training was conducted using four datasets about three plants, a public Arabidopsis dataset(CVPPP2014), and three developmental multi-view datasets from Arabidopsis, maize, and physalis. Multi-view images of given plants were collected at different developmental periods. The Arabidopsis datasets contained only leaf, and the maize and physalis datasets were different from the Arabidopsis datasets, having clearly distinct leaf, stems, and petioles. The results showed that the mean accuracy precision(mAP0.5) of the Mask-RCNN model for Arabidopsis in the public datasets which was in the same growth period reached 85.3% and the mean intersection over union(mIOU) was 73.4%. The mean accuracy precision was more than 70.0% across different growth periods of Arabidopsis, maize, and physalis. The mean intersection over union was more than 60.0% across different growth periods of Arabidopsis, which indicated that Mask-RCNN displayed satisfying versatility for plant phenotyping and had high value for plant phenotyping. The results showed that the model had competitive advantage compared with previous plant segmentation algorithms. Furthermore, taking advantage of multi-view images, a leaf tracking method was presented to solve the problem of plant occlusions. It was helpful for the leaf counting and leaf area calculation of plants. The results showed that the proposed methods had a superior performance compared with other existing plant segmentation algorithms, and was promising to build a dynamic modeling for various plants during their entire growth cycles.

     

/

返回文章
返回