Ni Xueping, Li Changying, Jiang Huanyu, Takeda Fumiomi
College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou, China.
College of Engineering, University of Georgia, Athens, GA USA.
Hortic Res. 2020 Jul 1;7:110. doi: 10.1038/s41438-020-0323-3. eCollection 2020.
Fruit traits such as cluster compactness, fruit maturity, and berry number per clusters are important to blueberry breeders and producers for making informed decisions about genotype selection related to yield traits and harvestability as well as for plant management. The goal of this study was to develop a data processing pipeline to count berries, to measure maturity, and to evaluate compactness (cluster tightness) automatically using a deep learning image segmentation method for four southern highbush blueberry cultivars ('Emerald', 'Farthing', 'Meadowlark', and 'Star'). An iterative annotation strategy was developed to label images that reduced the annotation time. A Mask R-CNN model was trained and tested to detect and segment individual blueberries with respect to maturity. The mean average precision for the validation and test dataset was 78.3% and 71.6% under 0.5 intersection over union (IOU) threshold, and the corresponding mask accuracy was 90.6% and 90.4%, respectively. Linear regression of the detected berry number and the ground truth showed an value of 0.886 with a root mean square error (RMSE) of 1.484. Analysis of the traits collected from the four cultivars indicated that 'Star' had the fewest berries per clusters, 'Farthing' had the least mature fruit in mid-April, 'Farthing' had the most compact clusters, and 'Meadowlark' had the loosest clusters. The deep learning image segmentation technique developed in this study is efficient for detecting and segmenting blueberry fruit, for extracting traits of interests related to machine harvestability, and for monitoring blueberry fruit development.
果实性状，如簇生紧密度、果实成熟度和每簇浆果数量，对于蓝莓育种者和生产者做出与产量性状、可收获性以及植株管理相关的基因型选择的明智决策非常重要。本研究的目的是开发一种数据处理管道，使用深度学习图像分割方法自动对四个南方高丛蓝莓品种（“翡翠”、“法辛”、“云雀”和“明星”）的浆果进行计数、测量成熟度并评估紧密度（簇紧密度）。开发了一种迭代注释策略来标记图像，从而减少注释时间。训练并测试了一个Mask R-CNN模型，以检测和分割单个蓝莓的成熟度。在0.5的交并比（IOU）阈值下，验证数据集和测试数据集的平均精度分别为78.3%和71.6%，相应的掩码准确率分别为90.6%和90.4%。检测到的浆果数量与真实值的线性回归显示，R值为0.886，均方根误差（RMSE）为1.484。对从四个品种收集的性状分析表明，“明星”每簇的浆果数量最少，“法辛”在4月中旬的成熟果实最少，“法辛”的簇最紧凑，“云雀”的簇最松散。本研究中开发的深度学习图像分割技术对于检测和分割蓝莓果实、提取与机械可收获性相关的感兴趣性状以及监测蓝莓果实发育是有效的。
