基于改进YOLO11的番茄叶片病害检测方法OA
An improved YOLO11-based method of detecting tomato leaf diseases
为解决番茄叶片病害检测中存在的类别分布不均、小目标特征提取困难及复杂环境检测精度不足等问题,提出了一种融合多模块的YOLO11改进算法.模型首先采用自适应阈值焦点损失函数(adaptive thresh-old focal loss,ATFL)替换传统的交叉熵损失,通过动态权重分配策略,针对病害样本不平衡问题优化训练过程,增强模型对低频病害类别的特征学习能力,提高整体检测精度;其次,引入混合池化注意力模块(hybrid pooling attention,HPA)替换C3k2模块中的普通卷积,设计并行池化分支,结合分组特征重校准技术,强化小目标病害的空间特征提取能力,提升复杂背景下的抗遮挡检测能力;最后,将Conv结构中的归一化替换为重参数批归一化(re-parameterized batch normalization,RepBN),引入可学习调节参数重构归一化过程,稳定训练阶段的特征分布,加速模型收敛并增强泛化性能.结果显示,改进后的YOLO11模型在番茄叶片病害数据集上有着很好的性能,其精确率、召回率、mAP@0.5和mAP@0.5:0.95评价指标相较于YOLO11基础模型分别提升了5.8、4.3、2.3和1.2百分点,特别是在小目标病害区域和密集遮挡场景下,具有很好的检测效果.结果表明,YOLO11改进算法有效提升模型对番茄叶片病害的检测性能,实现番茄叶片病害检测精度与鲁棒性的同步提升.
An improved YOLO11 algorithm integrating multiple modules was proposed to solve the problems of uneven distribution of categories,difficulties in extracting features of small target,and insuffi-cient accuracy in detecting tomato leaf diseases in complex environments.Firstly,the adaptive threshold fo-cal loss(ATFL)was used to replace the loss of traditional cross-entropy.The process of training was opti-mized by implementing a dynamic weight allocation strategy to address the issue of imbalanced samples of diseases,enhance the ability of model to learn features for disease categories with low-frequency and im-prove the overall accuracy of detection.Secondly,the hybrid pooling attention module(HPA)was intro-duced to replace the ordinary convolution in the C3k2 module,and a parallel pooling branch was designed.The ability of extracting spatial feature of small target diseases was strengthened by combining with the group feature recalibration technique to enhance the ability of antiocclusion detection in complex environ-ments.Finally,the normalization in the Conv structure was replaced with re-parameterized batch normaliza-tion(RepBN),and a learnable adjustment parameter was introduced to reconstruct the process of normal-ization to stabilize the feature distribution at the stage of training,accelerate the convergence of model,and enhance the performance of generalization.The results showed that the improved YOLO11 model had excel-lent performance on the dataset of tomato leaf diseases,with increases of 5.8,4.3,2.3,and 1.2 percent-age points in the precision,recall,mAP@0.5,and mAP@0.5:0.95 evaluation metrics compared with that of the YOLO11 base model,respectively.It had good performance on detecting areas of small target diseas-es and scenarios of dense occlusion in particular.It is indicated that the improved YOLO11 algorithm effec-tively improves the performance of the model in detecting tomato leaf diseases,achieving a synchronous im-provement in the accuracy and robustness of detecting tomato leaf diseases.
秦皓翔;赵霞;王敏;周慧
甘肃农业大学信息科学技术学院,兰州 730070甘肃农业大学信息科学技术学院,兰州 730070甘肃农业大学信息科学技术学院,兰州 730070甘肃农业大学信息科学技术学院,兰州 730070
农业科技
深度学习YOLO11番茄叶片病害检测损失函数HPARepBN
deep learningYOLO11detection of tomato leaf diseasesloss functionhybrid pooling attention module(HPA)re-parameterized batch normalization(RepBN)
《华中农业大学学报》 2026 (3)
115-126,12
甘肃省自然科学基金项目(24JRRA656)
评论