Yao Na, Ni Fuchuan, Wu Minghao, Wang Haiyan, Li Guoliang, Sung Wing-Kin
College of Informatics, Huazhong Agricultural University, Wuhan, China.
Hubei Engineering Technology Research Center of Agricultural Big Data, Wuhan, China.
Front Plant Sci. 2022 May 25;13:876357. doi: 10.3389/fpls.2022.876357. eCollection 2022.
Peach diseases seriously affect peach yield and people's health. The precise identification of peach diseases and the segmentation of the diseased areas can provide the basis for disease control and treatment. However, the complex background and imbalanced samples bring certain challenges to the segmentation and recognition of lesion area, and the hard samples and imbalance samples can lead to a decline in classification of foreground class and background class. In this paper we applied deep network models (Mask R-CNN and Mask Scoring R-CNN) for segmentation and recognition of peach diseases. Mask R-CNN and Mask Scoring R-CNN are classic instance segmentation models. Using instance segmentation model can obtain the disease names, disease location and disease segmentation, and the foreground area is the basic feature for next segmentation. Focal Loss can solve the problems caused by difficult samples and imbalance samples, and was used for this dataset to improve segmentation accuracy. Experimental results show that Mask Scoring R-CNN with Focal Loss function can improve recognition rate and segmentation accuracy comparing to Mask Scoring R-CNN with CE loss or comparing to Mask R-CNN. When ResNet50 is used as the backbone network based on Mask R-CNN, the segmentation accuracy of segm_mAP_50 increased from 0.236 to 0.254. When ResNetx101 is used as the backbone network, the segmentation accuracy of segm_mAP_50 increased from 0.452 to 0.463. In summary, this paper used Focal Loss on Mask R-CNN and Mask Scoring R-CNN to generate better mAP of segmentation and output more detailed information about peach diseases.
桃树病害严重影响桃子产量和人们的健康。精确识别桃树病害并分割病害区域可为病害防治和治疗提供依据。然而，复杂的背景和不均衡的样本给病斑区域的分割和识别带来了一定挑战，且困难样本和不均衡样本会导致前景类和背景类的分类准确率下降。在本文中，我们应用深度网络模型（Mask R-CNN和Mask Scoring R-CNN）对桃树病害进行分割和识别。Mask R-CNN和Mask Scoring R-CNN是经典的实例分割模型。使用实例分割模型可以获得病害名称、病害位置和病害分割结果，且前景区域是下一次分割的基本特征。焦点损失（Focal Loss）可以解决由困难样本和不均衡样本引起的问题，本文将其用于该数据集以提高分割精度。实验结果表明，与使用交叉熵损失（CE loss）的Mask Scoring R-CNN或Mask R-CNN相比，带有焦点损失函数的Mask Scoring R-CNN可以提高识别率和分割精度。当基于Mask R-CNN使用ResNet50作为骨干网络时，segm_mAP_50的分割精度从0.236提高到0.254。当使用ResNetx101作为骨干网络时，segm_mAP_50的分割精度从0.452提高到0.463。综上所述，本文在Mask R-CNN和Mask Scoring R-CNN上使用焦点损失以生成更好地分割平均精度均值（mAP）并输出有关桃树病害更详细的信息。
