基于改进Mask R-CNN模型的植物叶片分割方法

通过图像处理对植物叶片进行分割是研究植物表型性状的基础,但叶片间相互遮挡、叶片边缘特征不明显以及幼叶目标过小会对叶片分割效果造成很大的障碍。针对上述问题,该研究提出了一种基于改进Mask R-CNN模型的植物叶片分割方法,通过引入级联检测模块对模型检测分支进行改进,以提高遮挡叶片检测质量;利用注意力机制和一个2层3×3卷积模块对模型分割分支进行优化,以提高边缘特征表达能力;在模型测试过程中采用多尺度叶片分割策略,利用多个尺度上的最优目标以分割幼叶。测试结果表明,在CVPPP叶片分割挑战数据集中的对称最佳Dice得分Symmetric Best Dice（SBD）为90.3%,相比于利用域随机化数据增强策略并采用Mask R-CNN模型进行叶片实例分割的方法提高了2.3个百分点,叶片分割效果有显著提升。该研究提出的方法可以有效解决植物叶片分割效果不佳的问题,为植物表型研究提供技术支撑。

Segmentation method for plant leaves using an improved Mask R-CNN model

Extracting the leaves has been one of the most fundamental researches for the phenotypic traits of plants. However, it is difficult to further improve the accuracy of leaves segmentation, due to the mutual occlusion between leaves in rosette plants. It is still lacking in the leaf edge characteristics, particularly for the small target of young leaves. In this study, an improved instance segmentation benchmark Mask R-CNN model was proposed to realize an accurate segmentation for the plant leaves. The Cascade R-CNN was also introduced to generate the selection of the proposal boxes for the region proposal network. The boxes were first sent into the Head1 convolutional network with a threshold of 0.5, and then the detection was input into the Head 2 convolutional network with a threshold of 0.6, and finally, the detection was fed into the Head 3 convolutional network with a threshold of 0.7. Three cascaded networks were used to gradually increase the threshold, where the output of the previous network was applied to the higher threshold of the next network, indicating the high quality of the detection branch. After the segmentation branch full convolutional layer (FCN), an attention mechanism (SeNet) and a two-layer 3×3 convolution module were added to weigh the features of the segmentation branch, and further to extract the leaf edge segmentation. The segmentation of multi-scale leaves was also adopted for the leaves of different sizes in an image in the test process. The original and enlarged images were input into the model at the same time for the leaves segmentation. The optimal target on multiple scales was achieved using Non-Maximum Suppression (NMS). The ResNeXt101 was selected as the backbone network to extract the characteristics of leaves. An improved model of deep learning was also utilized at the same time. The image enhancement included random mirroring, rotation, enlargement, and size adjustment. There was no longer requiring for the special function library support, indicating the simplified data enhancement. The improved model had effectively reduced the blade occlusion, edge blur, and small target in the blade segmentation. Correspondingly, the Symmetric Best Dice (SBD) was 90.3% in the CVPPP Leaf Segmentation Challenge (LSC), which was 2.3 percentage points higher than that in Mask R-CNN, particularly than 1 percentage points than before. The data enhancement strategy was greatly contributed to improving the neural network model suitable for the leaf segmentation. The finding can provide a more accurate leaf segmentation for the plant phenotype using image processing, indicating a high application value in the research field of plant leaves.