The main challenge in deep learning related to the identification of grape leaf diseases is how to achieve good performance in the case of available sparse datasets or limited number of annotated samples, small lesions, redundant information and blurred background information in grape leaf disease images. This paper proposes a three-stage deep learning-based pipeline, including a convolutional neural netword (Faster R-CNN) for detection of lesions, a generative adversarial network (DCGAN) for data augmentation and a residual neural network (ResNet) for identification of lesions, to solve these problems. Firstly, Faster R-CNN was used to mark the location of grape leaf lesions to obtain lesions dataset for data augmentation and ResNet for identification of lesions. Secondly, leaf lesion images were fed into DCGAN to generate synthetic grape lesion images for identification of lesions. Finally, ResNet trained in the training dataset consisting of real grape leaf lesions and synthetic grape lesion images obtained by DCGAN, was used to identify the grape leaf lesions according to the majority voting principle. The first experimental results showed that Faster R-CNN, DCGAN and ResNet had good performance in each stage. Secondly, the second experimental results showed that the proposed three-stage method is superior to single-stage and two-stage methods in the case of sparse datasets and small lesions. Finally, a generalization experiment was also carried out using 100 images on the internet to verify the generalization of the proposed three-stage method. Experimental results showed that the proposed three-stage method had good generalization ability.

与葡萄叶病识别相关的深度学习面临的主要挑战是如何在可用的数据集稀疏或注释样本数量有限、病斑小、信息冗余和葡萄叶病图像背景信息模糊的情况下取得良好的性能。本文提出了一个基于深度学习的三阶段管道，包括用于检测病变的卷积神经网络（Faster R-CNN）、用于数据增强的生成对抗网络（DCGAN）和用于识别病变的残差神经网络（ResNet）。病变，来解决这些问题。首先，使用 Faster R-CNN 标记葡萄叶片病变的位置，获取病变数据集进行数据增强，并使用 ResNet 进行病变识别。第二，叶子病变图像被输入 DCGAN 以生成合成葡萄病变图像以识别病变。最后，在由DCGAN获得的真实葡萄叶片病变和合成葡萄病变图像组成的训练数据集中训练的ResNet，用于根据多数投票原则识别葡萄叶片病变。第一个实验结果表明，Faster R-CNN、DCGAN和ResNet在每个阶段都有很好的表现。其次，第二个实验结果表明，在稀疏数据集和小病灶的情况下，所提出的三阶段方法优于单阶段和两阶段方法。最后，还使用互联网上的 100 张图像进行了泛化实验，以验证所提出的三阶段方法的泛化性。