适用于小样本显微图像数据集的柑橘黄龙病快速诊断模型

为了探究柑橘黄龙病病原菌对宿主叶片主脉显微结构的影响并建立基于叶片主脉显微图像的快速诊断方法，该研究以健康、染病未显症、染病显症和缺镁四类柑橘叶片主叶脉的显微图像为研究对象，提出了一个适用于小样本显微图像数据集的增强特征的无监督训练柑橘黄龙病检测模型（Enhanced Huanglongbing Unsupervised Pre-training Detect Transformer，E-HLBUP-DETR）。该模型首先采用无监督训练结合迁移学习构成上游网络（unsupervised pre-training model），再利用Yolact模型设计出增强特征网络（Enhanced Feature Network，EFN）与DETR(Detect Transformer)相结合构成下游网络，最终建立E-HLBUP-DETR诊断模型。研究结果表明，E-HLBUP-DETR模型检测的准确率可达96.2%，能够解决采用小规模数据集训练的模型存在过拟合和准确率低的问题。相较于未改进的DETR模型，E-HLBUP-DETR具有更高的检测准确率，识别准确率也优于CNN架构ResNext的92.1%与MobileNet的76.3%。研究结果可为显微尺度下柑橘黄龙病的早期快速诊断提供技术支持。

Detecting citrus Huanglongbing from few-shot microscopic images using an improved DETR

Abstract: Citrus Huanglongbing (HLB) is one of the most devastating diseases in the citrus industry. The HLB disease is often caused by the phloem-restricted, non-culturable gram-negative bacteria, Candidatus Liberibacter asiaticus (Las). It is very necessary to rapidly and accurately diagnose the infected citrus for the prevention of the disease so far. In this study, an unsupervised training model with the enhanced feature network for the few-shot microscopic images (Enhanced Huanglongbing Unsupervised Pre-training Detect Transformer, E-HLBUP-DETR) was proposed to investigate the effect of Las on the microstructure of the main veins of the citrus leaves, particularly for the rapid detection of HLB disease. The model was mainly composed of the upstream and downstream networks. Specifically, the transformer and unsupervised training were used in the upstream network. A pre-trained model was then generated to train on the ImageNet for the downstream detection. Two components were divided in the downstream network, including the enhanced feature network and detect transformer (DETR) network. A Yolact model was also designed in the enhanced feature network to extract the features of the region where the Las located. The ResNet was selected as the backbone network, where the local coefficient generation branch was added into two existing parallel branches in the feature extraction network of Yolact model. The resulting local coefficients were then used to weight the local mask in the Protonet for the more accurate bounding box of region of interests (phloem, xylem and pith) by the fast non-maximum suppression (NMS), which was used to enhance the localization of feature region. The enhance feature network was combined the region of interest with the original image for the better scale of dataset. The enhanced dataset was finally fed into the DETR, in order to realize the combination of DETR and enhanced feature network, which was called as the downstream model. The upstream and downstream networks were then assembled the model (called as Enhanced Huanglongbing Unsupervised Pre-training Detect Transformer (E-HLBUP-DETR)). The results showed that an excellent generalization ability of the improved model was achieved with the average precision and recall more than 99%. Among them, the detection accuracy of E-HLBUP-DETR model was reached 96.2%, particularly with the high accuracy, but without the overfitting caused by the few-shot datasets. The improved DETR model also presented a much higher recognition accuracy than the original ones. Additionally, two representative convolutional neural networks (CNNs) were also introduced to detect the HLB at the micro scale, including the ResNeXt (a deep CNN) and MobileNet (a lightweight CNN). A better performance of E-HLBUP-DETR model was achieved with the detecting accuracies of 92.1% and 76.3%, compared with the ResNeXt and MobileNet, respectively. Therefore, the E-HLBUP-DETR model can also be expected to serve as a much larger receptive field for the excellent performance under a large amount of training data and parameters. The finding can provide technical supports for the detection of citrus Huanglongbing during citrus production.