华南农业大学学报2025，Vol.46Issue(3)：399-406,8.DOI:10.7671/j.issn.1001-411X.202410019
基于特征交互的样本不均衡的玉米病害检测方法
A maize disease detection method based on feature interaction under imbalanced sample condition
摘要
Abstract
[Objective]To address the issues of imbalanced data samples and low detection accuracy in maize leaf disease detection under complex environments.[Method]An improved object detection network SF_YOLOv5 was proposed.First,based on the multi-scale pyramid structure of YOLOv5,a novel spatial-feature pyramid structure(SPD-FPN)was designed to enhance the network's ability to recognize small-target disease features at high-resolution levels while retaining large-target information at low-resolution levels,thereby improving overall detection accuracy and robustness.Second,the Focal Loss function was introduced to increase the weight of hard-to-classify samples and reduce the influence of easily classified samples,ensuring that the model focused more on the minority samples often overlooked in imbalanced datasets.Additionally,transfer learning was applied to the design of SF_YOLOv5,where pre-trained YOLOv5 model parameters were transferred to the improved SF_YOLOv5 network for training.This leveraged knowledge from large-scale datasets to enhance the model's generalization capability for maize disease detection.[Result]Experimental validation on the constructed maize disease dataset showed that SF_YOLOv5 achieved a mean average precision(mAP)of 93.3%and a recall of 89.6%,significantly outperforming the original YOLOv5 model.And the model was small in size,and had been deployed in mobile devices.[Conclusion]The results demonstrate that the improved network performs better than the original model in detecting maize leaf diseases under imbalanced sample conditions.This approach can be applied to intelligent diagnosis of maize diseases in farmland scenarios with imbalanced data,providing a theoretical foundation for real-time maize disease detection in the agricultural sector.关键词
深度学习/玉米病虫害/特征交互/样本不均衡Key words
Deep learning/Maize pest and disease/Feature interaction/Imbalanced sample分类
农业科技引用本文复制引用
姜飞,叶炜,李兆星,王洪凯,王教瑜..基于特征交互的样本不均衡的玉米病害检测方法[J].华南农业大学学报,2025,46(3):399-406,8.基金项目
浙江省"尖兵"和"领雁"研发攻关计划(2023C02018) （2023C02018）
浙江省湖州市科技局农业"双强"专项(2023ZD2039) （2023ZD2039）
