Abstract Effective segmentation of plant leaves is very necessary for non-contact extraction of plant leaf phenotype, especially leaf phenotype under environmental stress. However, the phenotype of leaves will change due to the influence of the environment, which increases the difficulty of detection. In this study, we proposed an accurate automatic segmentation method that combines Mask R-CNN with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm based on RGB-D camera to segment overlapped poplar seedling leaves under heavy metal stress. Firstly, an effective encoding method of depth information was used to facilitate the feature extraction of depth information. Next, we deployed Mask R-CNN to train the RGB-D data and fuse their features in the FPN structure to obtain more accurate leaf areas. Based on the detected leaf areas and depth data, DBSCAN based on manifold distance was then applied to segment a single leaves from overlapping leaves in the detected areas. Several analyses were performed to evaluate the performance of the proposed method, including the comparison of our network with classic Mask R-CNN and the comparison of DBSCAN based on manifold distance with other classic clustering methods. We used the pixel-wise Intersection over Union (p-IoU) to evaluate the detection results more accurately. In the experiments, the obtained p-IoU of normal and stressed leaves was 0.885 and 0.874, respectively, with corresponding mean accuracy values of 0.897 and 0.888. From our experimental results, it can be concluded that the proposed method can automatically detect leaves with high accuracy, which can be applied to 3-D leaf phenotype research and automatic plant de-leafing.

中文翻译：

---

结合Mask R-CNN和DBSCAN的重叠杨树苗叶自动分割

摘要 植物叶片的有效分割对于非接触式提取植物叶片表型，尤其是环境胁迫下的叶片表型非常必要。但是，叶片的表型会因环境的影响而发生变化，增加了检测的难度。在这项研究中，我们提出了一种准确的自动分割方法，该方法将 Mask R-CNN 与基于 RGB-D 相机的基于密度的应用程序空间聚类（DBSCAN）聚类算法相结合，以分割重金属胁迫下重叠的杨树幼苗叶片。首先，采用有效的深度信息编码方法，方便深度信息的特征提取。接下来，我们部署了 Mask R-CNN 来训练 RGB-D 数据并将它们的特征融合到 FPN 结构中以获得更准确的叶子区域。基于检测到的叶子区域和深度数据，然后应用基于流形距离的 DBSCAN 从检测区域的重叠叶子中分割出单个叶子。进行了多项分析以评估所提出方法的性能，包括我们的网络与经典 Mask R-CNN 的比较以及基于流形距离的 DBSCAN 与其他经典聚类方法的比较。我们使用像素级的联合交集（p-IoU）来更准确地评估检测结果。在实验中，获得的正常叶和受压叶的 p-IoU 分别为 0.885 和 0.874，相应的平均准确度值为 0.897 和 0.888。从我们的实验结果可以得出结论，所提出的方法可以高精度地自动检测叶子，