Precise and rapid detection is crucial to early warning and timely control of the crop leaf diseases in agricultural production. However, the conventional disease detection has limited to the low efficiency, high cost, and strong subjectivity, difficult to meet the requirements of modern agriculture in recent years. Alternatively, deep learning technologies can be expected to effectively address these limitations, particularly the convolutional neural networks (CNN), due to their powerful feature extraction and automatic learning. The detection efficiency and accuracy can be significantly enhanced to emerge as a hot topics in this field. In datasets, the existing resources are primarily categorized into the institutionally constructed datasets and public repositories. Furthermore, the research scope can be classified into single- and multi-crop datasets. High-quality datasets can be characterized by three important indicators: the large-scale data volume, diversity of samples, and balanced class distribution.This study aims to systematically review the applications of CNN in crop leaf disease detection from two perspectives: image classification and object detection. In image classification, the disease identification tasks were categorized into disease type classification, severity level assessment, and combined type-severity identification. The CNN algorithms were often combined with the optimization strategies, like the attention mechanism and transfer learning, indicating the highly efficient and accurate performance. Disease type identification was categorized into the single and complex background. Single background research was typically achieved in the high recognition accuracy but with the limited generalization. In contrast, the complex background studies focused mainly on the natural field environments, which were closer to the practical applications. While the lightweight model was prioritized to facilitate the deployment on edge devices. And the promising trends were advanced towards the high accuracy, anti-interference and model lightweight. Disease severity grading was achieved in the quantitative classification, according to the lesion characteristics, such as the area and distribution patterns. Many models also demonstrated the high accuracy, computational efficiency, and robustness against interference. The precise assessment of the disease severity was offered the decision-making on the targeted pesticide application. Simultaneously, the integrated recognition over different disease types and severity levels was emerged as a hot research direction. In object detection, the current research was focused on the representative two-stage algorithms (represented by Faster R-CNN and Mask R-CNN) and single-stage algorithms (represented by the SSD series and YOLO series). While two-stage algorithms was typically achieved the higher accuracy and robustness, thereby suffering from the large model sizes and slow inference speeds for the numerous region proposals. Additionally, there was the high dependency on dataset quality and scale. Consequently, the mainstream object detection models were required for the crop leaf disease detection in practical applications. Conversely, the single-stage algorithms shared the small model sizes and high recognition efficiency. But the lower precision and higher miss rates were also observed in complex scenarios. The reason was attributed to the candidate box set, in order to select the foreground part of the disease image. Notably, the YOLO series algorithms can be expected to serve as the mainstream models after continuous iterative optimization. There was the better balance between accuracy and efficiency, indicating the high practical value. Nevertheless, numerous challenges included the confusion caused by the similar lesion features, complex and variable field conditions, and the constraints of the limited datasets on model generalization. Future research should focus on the lightweight, high-precision models to overcome computational constraints in agricultural settings; Model generalization to construct the cross-regional/seasonal disease datasets spanning diverse spatiotemporal conditions; the robustness through multimodal data fusion (e.g., text, sensors, and thermal imaging); simultaneous multi-disease detection within single samples via multi-label learning; and the precise, intelligent disease control technologies for the complex field environments in the large-scale farmland. The finding can also provide a strong reference to facilitate the targeted pesticide application for the ecologically sustainable agriculture.

A method for measuring pork freshness based on images and the EfficientNet framework was established. A total of 2500 images of pork with different freshness were collected as original dataset and processed by image enhancement to construct a new dataset of 60000 images. First, EfficientNet was trained with the CIFAR-10 dataset to determine the basic structure and initial weights of the model. Then, the model was trained and improved using the constructed dataset to make the model suitable for five classification problems. Finally, the established model was tested, verified, and compared with the current mainstream convolutional neural network (CNN) models of Alexnet, VGG16 and ResNet50. The results showed that the average correct recognition rate of the EfficientNet model was as high as 98.62%, which was significantly better than that of the Alexnet, VGG16 and ResNet50 models. The correct recognition rate of the EfficientNetB2 model was 99.22%, and the training time was only 157 min. The comprehensive performance of the EfficientNetB2 model was the best, making it the most suitable method for pork freshness identification. In order to improve its generalization ability, the optimizer algorithm of the EfficientNetB2 model was improved, and the performances of stochastic gradient descent (SGD), adaptive moment estimation (Adam), root mean square propagation (RMSProp) and rectified adaptive moment estimation (RAdam) were compared. The results showed that the RAdam optimizer failed to further improve the accuracy of the model but instead helped to improve its generalization capability, which will of practical significance for engineering applications.