Abstract: In the process of crop production management, it is very important to diagnose crop diseases timely and accurately. Image recognition based on deep learning provided a new approach for automatic diagnosis of crop diseases. Compared with traditional pattern recognition methods which are used in image recognition, deep learning neural network models are able to extract features automatically, obtain high-dimensional features from low-dimensional features and achieve better learning results. This paper systematically reviewed the recent development of deep learning in automatic image recognition, and the concepts related to shallow networks were first introduced, on top of that, the advantages of deep learning methods were expounded, and the widely used image recognition algorithm, convolutional neural network was briefly introduced. Diagnosis of crop disease based on image recognition could be classified as diagnosis of a single disease in a single crop, multiple diseases in a single crop, and multiple diseases of various crops. After discussing and analyzing the research status of deep learning methods applied in these three areas, and the difficulties and challenges faced at present, prospective comments on the technology bottlenecks in this field 
that might be broken through in the future were put forward.

Key words: deep learning, image recognition, disease diagnosis, convolutional neural network

摘要： 在作物生产管理过程中，正确及时地诊断作物所患病害非常关键。基于深度学习的图像识别为作物病害自动快速诊断提供了新途径。相比传统图像识别所用的模式识别方法，深度学习网络模型能自行提取特征且能够由低维特征抽象出高维特征，取得更好的学习效果。系统梳理了深度学习在图像自动化识别方面的发展历程，介绍了浅层神经网络的相关概念，阐述了深度学习与之相比具有的优势，并简述了深度学习的重要图像识别算法——卷积神经网络。作物病害图像识别由单作物单病害、单作物多病害和多作物多病害的识别三部分组成，在分析讨论深度学习这三方面的研究现状以及目前该领域面临的困难与挑战的基础上，提出了未来可能突破的难点和研究重点。

关键词: 深度学习, 图像识别, 病害诊断, 卷积神经网络