基于图像和卷积神经网络的蝴蝶兰种苗生长势评估

在蝴蝶兰(Phalaenopsisaphrodite)产业中,种苗在达到最短营养栽培时长时的生长势在其后续的栽培链和最终的经济利润中起着重要的作用。当前在商业大型温室中主要采取人工方式对每株种苗进行评估,既费时又费力。基于RGB图像进行植物生长评估的相关研究依赖于从图像中手动提取人工定义的特征,从而影响了机器学习模型的有效性和泛化能力。该研究使用卷积神经网络(Convolutional Neural Network,CNN)来探讨其以端对端方式评估温室中蝴蝶兰种苗生长势的可行性。对在温室中采集的图像数据集,采用不同的CNN架构(VGG、ResNet和Inception-v3)结合不同的训练机制(从头训练、微调、特征提取)建立基准模型,其中微调取得了最佳的分类结果。考虑到该研究的目标任务是对具有复杂图像背景的单个温室种苗的形态分类,为进一步提高模型性能,在可控的实验室条件下采集了更多的种苗图像。实验室图像进行背景分割后,用于协助模型更好地学习植株的形态,即建立增强模型。与基准模型相比,2种增强方式总体上在温室测试集的F1-score取得了0.03~0.05的提升。采用增强方式Ⅱ的VGG模型取得了最高的性能(温室测试集上的F1-score为0.997),并对该模型的特征图进行可视化。在高层特征图中,目标种苗区域被激活,同时滤除了大部分背景(包括相邻种苗的叶片),进一步证明了能够采用CNN对温室种苗进行有效的形态学习和刻画。总体结果表明,深度学习模型可用于基于图像的蝴蝶兰种苗生长势评估,并且可扩展用于温室下其他植物类型的生长评估。

Image-based assessment of growth vigor for Phalaenopsis aphrodite seedlings using convolutional neural network

Abstract: In the Phalaenopsis industry, the growth vigor of seedlings when reaching their minimum growth time of vegetative cultivation plays an important role in the subsequent production chain and the final economic profits. The current manual assessment taking place in the commercial large-scale greenhouse is time-consuming and labor-intensive. Related studies based on RGB image for plant growth assessment relied on extracting hand-crafted features from images, affecting the effectiveness and generalization ability of machine learning models. In this study, the Convolutional Neural Network (CNN) was employed to explore its feasibility in assessing the growth vigor of Phalaenopsis aphrodite seedlings grown in the greenhouse in an end-to-end manner. Seedling images were collected in the greenhouse conditions with complex image background. Baseline models on the greenhouse dataset were established using different CNN architectures (VGG, ResNet, Inception-v3) coupled with various training mechanisms (training from scratch, fine-tuning, feature extraction), in which fine-tuning achieved the best classification results. Considering the target task of morphological classification for individual greenhouse seedlings with complex image background, to further boost model performance, additional seedlings images were acquired in controlled laboratory conditions. The segmented laboratory images were used to assist in model learning, namely building the augmented models. Two approaches were adopted, achieving an overall improvement in the testing F1-score of 0.03-0.05 compared with baseline models. The VGG model with augmentation method II achieved the highest performance in this study (F1-score of 0.997 on the test set), its feature maps were also visualized. In higher-level feature maps, regions of the target seedling were activated while filtering out most background including leaves from adjacent seedlings, proving the effective morphology characterizing using CNN for greenhouse seedlings. The overall results demonstrated the potential of deep learning models for image-based assessment of growth vigor for Phalaenopsis aphrodite seedlings and maybe other kinds of plants in greenhouse conditions.