AUTHOR=Yao Na , Ni Fuchuan , Wu Minghao , Wang Haiyan , Li Guoliang , Sung Wing-Kin TITLE=Deep Learning-Based Segmentation of Peach Diseases Using Convolutional Neural Network JOURNAL=Frontiers in Plant Science VOLUME=Volume 13 - 2022 YEAR=2022 URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.876357 DOI=10.3389/fpls.2022.876357 ISSN=1664-462X ABSTRACT=Peach diseases seriously affect their 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, but the complex background and imbalance samples bring certain challenges to the segmentation and recognition of the lesion area. This paper studies segmentation and recognition of peach diseases based on deep network models (Mask R-CNN and Mask Scoring R-CNN). Mask R-CNN and Mask Scoring R-CNN both includes two stages, one stage is feature extraction, the other stage is classification and detection box regression and Mask prediction. RPN (RegionProposal Network) belongs to the one stage which is for selecting targeted features. There are two branches in the RPN network. One branch is to classify by softmax whether the feature is a foreground class (positive) or a background class (negative). The foreground area is basis feature for next segmentation and so this step is important. But, the difficult-to-classify samples and imbalance samples can lead to a decline in classification of foreground class and background class. In the experiment of this paper, the lesion area of peach tree is a positive sample. If the lesion area of peach tree is small, then there will be few positive samples, which will result in an imbalance of positive and negative samples, and which may also make some positive samples become difficult samples. So, Focal Loss was used for this dataset to improve segmentation accuracy. This Loss function was used in RPN network.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 is increased from 0.236 to 0.254, and the bbox_mAP_50 is increased from 0.396 to 0.534. When ResNetx101 is used as the backbone network, the segmentation accuracy of segm_mAP_50 is increased from 0.452 to 0.463, and the bbox_mAP_50 is increased from 0.749 to 0.771.