Accurate detection and identification of crop diseases is an important measure to promote the development of intelligent and modernized agricultural production. With the development of computer vision technology, deep learning methods have been rapidly applied, and the use of convolutional neural networks to detect and identify crop diseases has become a hot research topic in recent years. In this paper, the disadvantages of traditional crop disease identification methods are analyzed. Based on the convolutional neural network model structure of crop disease detection and recognition, combined with the development and optimization process of convolutional neural network model, the specific application of convolutional neural network in crop disease detection and recognition is classified. From based on public data sets and selfbuilt data set of crop disease classification; based on double stage of target detection and single phase detection of crop disease detection; foreign and domestic crop disease severity evaluation of three aspects, the research progress of convolutional neural network model are summarized, while also contrasted its performance. The current problems of convolutional neural network model based on crop disease detection and recognition are pointed out: the network model with good recognition performance on public data sets has poor recognition performance on selfbuilt complex data sets. The crop disease detection algorithm based on twostage target detection has poor realtime performance and is not suitable for small target detection; The detection accuracy of crop disease based on singlestage target detection is low in complex background. The accuracy of crop disease degree assessment model in complex field environment is low. Finally, the future research directions are prospected as follows: how to obtain highquality crop disease data sets; how to improve the network generalization performance; how to improve crop monitoring performance in field environment; how to locate the area of large area crop disease, and to evaluate the severity of disease and early warning of single branch crop disease.

In order to achieve fast and accurate obstacle detection in vineyard field on unmanned agricultural machinery, four kinds of lightweight target detection neural networks, EfficientDetD0, YOLOV4TINY, YOLOV3TINY, and YOLOFASTEST, were trained with the independently built vineyard field scene dataset,and the training model eretransplanted to the edge device NVIDIA JETSON TX2. The accuracy of obstacle detection and the applicability of the four models on TX2 were evaluated. The results showed that the mAP of YOLOV3TINY, YOLOV4TINY, EfficientDetD0, and YOLOFASTEST were 0.648, 0.601, 0.598, and 0401, respectively. The experimental results on TX2 showed that the realtime video detection frame rates of the above network models were 34.24 frames, 24.75 frames, 2.34 frames, and 2.97 frames, respectively. Among the four target detection networks, YOLOV3TINY has the highest detection accuracy on the dataset and the fastest realtime detection speed, but it also consumes high hardware resources relatively. When considering the hardware resource consumption, YOLOV4TINY can better balance detection accuracy, running speed, and hardware resource consumption and achieve good results when running multiple tasks.

In order to obtain high accuracy of plant disease classification with only a few training samples, a few-shot learning model was used as the disease classifier, and five kinds of shallow networks, Conv4, Conv6, ResNet10, ResNet18, and ResNet34, were used as feature extraction networks under the framework of three typical few-shot learning algorithms, including MatchingNet, ProtoNet, and RelationNet. Their performances were compared on the plant disease data set of PlantVillage. Under the condition of 5way、1shot, the average accuracies of MatchingNet, ProtoNet, and RelationNet were 72.29%, 72.43%, and 69.45%, respectively. ProtoNet+ResNet34 was the optimal combinational mode, and the accuracy reached 77.60%. Under the condition of 5way、5shot, the average accuracies of MatchingNet, ProtoNet, and RelationNet were 87.11%, 87.50%, and 82.92%, respectively. The accuracies were significantly improved compared to that of the 1shot condition. ProtoNet+ResNet34 was still the optimal one with an accuracy of 89.66%. The above test results show that by optimizing the combination of a few-shot learning framework and feature extraction network, the recognition model can achieve good effects for diseases with a small number of samples.