An accurate and rapid detection is often required to monitor and prevent the field crop pests. However, the high miss rates and insufficient accuracy can also be caused by the small size and dense distribution of the field pests. In this study, a small object detection was proposed for the crop pest images using improved YOLOv8, named FCDM-YOLOv8. Firstly, the original C2f module was replaced with a lightweight C2f-FE module in the backbone network, in order to reduce the computational burden of the model. Additionally, the depthwise separable convolution (DWConv) was introduced to replace the ordinary convolutions in both the backbone and neck networks. Furthermore, the number of the parameters was reduced to effectively enhance the detection performance and operational efficiency. Secondly, a context aggregation network (context aggregation network, CONTAINER) was incorporated in the neck network. The contextual information was then strengthened to refine the feature representations. The accuracy of the detection was improved for the better capture and recognition of the dense pest groups. Thirdly, the model structure was adjusted to remove the P5 layer and the large object detection head in the backbone network. A small object detection layer was added after modifications. More feature information related to small object was retained to detect the pests of small sizes. Fourthly, the decoupled head in YOLOv8 was replaced with a dynamic detection head (dynamic head, and Dyhead). The dynamic detection head was adaptively adjusted the detection strategies, according to the density of object regions. The dense and small objects were effectively focused to extract the more useful feature information. Finally, Focaler-MPDIoU was selected as the bounding box loss function, in order to improve the detection accuracy and robustness on the small objects and difficult examples. An experiment was also carried out to validate the improved model. The result show that the FCDM-YOLOv8 model was achieved in the precision, recall, mAP_0.5, and mAP_0.5~0.95 of 81.4%, 73.5%, 80.1%, and 41.1%, respectively, on the self-collected and constructed dataset of the field environment pest. Compared with the baseline YOLOv8n, the FCDM-YOLOv8 model was improved precision by 2.0 percentage points, recall by 5.2 percentage points, mAP_0.5 by 5.1 percentage points, and mAP_0.5~0.95 by 2.8 percentage points. Additionally, the model size was reduced by 38.1%. Compared with the mainstream object detection (such as Faster R-CNN, SSD, and other YOLO series models), the FCDM-YOLOv8 model demonstrated the highest recall rate and mAP values, with the lowest memory footprint. Visual comparisons with the baseline model also showed that the FCDM-YOLOv8 model was significantly improved the detection accuracy for the less miss rates. Furthermore, the generalization experiments were conducted on the public dataset VisDrone2019. The precision, recall rate, mAP_0.5, and mAP_0.5~0.95 of the FCDM-YOLOv8 model reached 52.6%, 38.9%, 41.1%, and 24.3%, respectively, which were 7.7, 5.3, 7.7, and 4.9 percentage points higher than the baseline YOLOv8n. On the dataset COCO2017-small, the precision, recall rate, mAP_0.5, and mAP_0.5~0.95 of the FCDM-YOLOv8 model reached 44.8%, 29.0%, 28.3%, and 16.0%, respectively, which were 1.9, 1.6, 2.2, and 1.5 percentage points higher than the baseline model YOLOv8n. The FCDM-YOLOv8 model shared the outstanding generalization and detection accuracy. Finally, we developed a small target detection system for crop pests based on the FCDM-YOLOv8 model. The system deployed the FCDM-YOLOv8 model at the back end and integrated the PyQt5 framework at the front end. It can accurately identify and locate wheat spiders and aphids, providing technical support for precision pesticide application. In addition, the system can count the number of targets in each image to evaluate pest density. In summary, this research provides technical support for the intelligent detection of small targets of crop pests in field environments.

Wheat diseases have posed a severe threat to the grain quality and yield in wheat production. It is highly required for the automatic, rapid, and accurate identification of wheat diseases on equipment with limited resources, particularly for timely preventive measures. In this study, a lightweight detection was proposed for wheat diseases using an improved YOLOv8n, termed PGCW-YOLOv8. Firstly, a lightweight CPU network (PP-LCNet) was introduced to replace the backbone network of the YOLOv8 structure, in order to reduce the large model weight files. Depthwise separable convolution (DepthSepConv) structure was introduced into the Backbone layer to reduce the parameter quantity, thereby reducing the size of weight files for better detection performance. Secondly, a global attention mechanism (GAM) module was added to the Neck section, in order to enhance the feature extraction and fusion of the network. As such, the model was improved to better focus on the small features of the target disease. The higher detection accuracy of minor lesions was achieved to better understand the important information in the image through a global attention mechanism, thereby accurately identifying the diseases under complex background and lighting conditions. Thirdly, a lightweight content-aware reassembly of features (CARAFE) module was introduced to aggregate the context information within a larger receptive field, in order to improve the detection accuracy of the model. The detailed image information was then effectively preserved using upsampling and downsampling operations. Finally, the Wise-IoU boundary loss function was used instead of the original loss function, in order to enhance the bounding box regression performance of the network model. The position and size of disease features were better learnt to improve the detection of small target diseases. Experimental results show that the improved PGCW-YOLOv8 model reduced the computational complexity (GFLOPs), parameters, and model size by 13.6%, 12.5%, and 11.3%, respectively, in the wheat disease datasets collected in field environments, compared with the original YOLOv8n baseline model. Meanwhile, the precision, mean average precision (mAP), and frames per second (FPS) of the improved model increased by 4.5, 1.9 percentage points, and 23.1%, respectively, compared with the original. A comparison was made under the same experimental conditions with the mainstream deep learning models, such as Faster R-CNN, YOLOv5s, YOLOv7, Yolov7-tiny, YOLOXs, and Edge-YOLO. The improved PGCW-YOLOv8 model shared the highest precision and mAP values, while the lowest computational complexity (GFLOPs), parameters, and model size. Three datasets of comparative experiments indicate that the lightweight PGCW-YOLOv8 network model outperformed the original YOLOv8 and YOLOv5 models, especially in the accurate detection of the wheat disease features under complex background and multi-target situations. The finding can provide a strong reference for the intelligent detection of wheat diseases in real time, particularly for rapid detection applications, such as deployment on unmanned aerial vehicles and mobile terminal equipment.