文献类型： 会议论文

第一作者： Zedong Shi

作者： Zedong Shi ¹ ; Jiahua Liu ¹ ; Xubin Qin ¹ ; Jing Cao ² ;

作者机构： 1.Tongda College of Nanjing University of Posts and Telecommunications, Yangzhou, China

2.Wuxi Branch of Jiangsu Academy of Agricultural Sciences, Wuxi, China

关键词： YOLO;Interpolation;Accuracy;Prevention and mitigation;Interference;Feature extraction;Inference algorithms

会议名称： Chinese Control Conference

主办单位：

页码： 8703-8708

摘要： A modified algorithm based on YOLOv5s is proposed for tomato leaf disease detection under interference from complex environmental factors and diverse types of diseases. Firstly, multiple disease images are preprocessed to enhance the richness of the dataset. In order to enhance the image classification ability of the model, the attention mechanism is introduced in the backbone network; the C 2 f module of YOLOv8 replaces the original C 3 module to obtain richer gradient flow information; the loss function SIoU is replaced to accelerate the convergence speed and improve the inference effect; CARAFE (ContentAware ReAssembly of FEatures) is utilized instead of the original nearest-neighbor interpolation method to expand the improved models receptive field, obtaining more feature map information effectively, thereby enhancing detection capabilities while ensuring lightweight design. In experiments, the improved algorithm achieves a $mathrm{mAP}_{0.5}$ of $93.3 %$ in test validation, outperforming the main models, such as the original YOLOv5s, YOLOv8, et al. The proposed tomato leaf disease detection method maintains good detection speed and high accuracy in complex real-world scenarios. This improved algorithm meets the requirements for precision recognition, providing theoretical support for the prevention and control of tomato leaf diseases.

分类号： tp273