1. 基于改进 YOLOv5 的柑橘病虫害检测.
- Author
-
李吴洁, 危疆树, 王玉超, 陈金荣, and 罗好
- Subjects
- *
LEAF diseases & pests , *LEAFMINERS , *CITRUS , *DATA mining , *ANTHRACNOSE - Abstract
[Objectives] After being infected by pathogens or pests, citrus leaves can lead to abnormal growth and development, reduced yield, and even death of citrus trees. Early detection of citrus leaf diseases and pests can effectively allow for preventive measures to reduce losses [Methods] In the actual detection process, the YOLOv5s model had problems such as inaccurate positioning and complex background. Inspired by the VAN(visual attention network)model, the LKA(large kernel attention)module was introduced to improve the YOLOv5s model. The improved YOLOv5s model could achieve centralized attention and fine extraction of image information; replacing conventional up sampling methods with CARAFE lightweight operators could improve feature reconstruction quality, solve scale mismatch problems, and enhance detection performance; using the FReLU activation function could capture more key features of citrus pests and diseases, improving detection accuracy. In addition, a dataset of citrus leaves containing anthracnose, ulcer disease, and infestation by leaf miner pests was constructed for experimentation. [Results] The improved model YOLOv5-LC showed the detection results of citrus pests and diseases as follows: the average detection accuracy mAP50 reached 94.5% and mAP50:95 was 84.3%, which were 2.0% and 4.4% higher than the original model, and the model size was only 7.3 MB, with 93.8% accuracy and 84.5% recall, and the number of floating-point operations was only 18.5 G. [Conclusions] The improved model YOLOv5-LC could more accurately detect citrus pests and diseases. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF