基于改进YOLOv9的番茄病害检测-《计算机技术与发展》

文章信息/Info

Author(s):: YAO Bai-wei; SHAO Jia-hui; XU Ying-jie; TIAN Hong; School of Rail Intelligent Engineering,Dalian Jiaotong University,Dalian 116000,China

Keywords:: YOLOv9; tomato disease detection; deep learning; attention mechanism; deformable convolution

摘要:: 在农业生产中,番茄病害的及时检测能够保护农作物的健康并且减少经济损失。番茄所处背景通常是复杂的,为了实现复杂背景下的番茄病害检测,提出了一种基于注意力机制的 YOLOv9 改进模型 TODC-YOLOv9。首先,引入了 DCNv3 可变形卷积对原始的 RepNCSPELAN4 模块进行改进后得到 DCNRCE 模块,实现在空间上的自适应聚合,从而增强模型对几何变换的适应能力和特征表示能力。其次,通过引入 SimAM 注意力机制进一步加强主干的特征提取能力,提高网络对输入数据的敏感度和识别能力,使模型能够支持更复杂的视觉任务。实验结果表明,TODC-YOLOv9 模型的精确率达到 98. 02% ,召回率达到 93. 75% ,平均精度均值 mAP_0. 5 和 mAP_0. 5:0. 95 分别达到 98. 27% 和 88. 72% ,均优于其他模型。由此表明,TODC-YOLOv9 模型具有较好的鲁棒性和识别能力,能够为复杂背景下的番茄病害检测提供有力的技术支持。

Abstract:: In agricultural production,timely detection of tomato diseases can protect the health of crops and reduce economic losses. The background of tomatoes is usually complex. In order to achieve tomato disease detection in complex backgrounds, we propose an improved YOLOv9 model TODC- YOLOv9 based on attention mechanism. Firstly,DCNv3 deformable convolution is introduced to improve the original RepNCSSPELAN4 module and obtain the DCNRCE module,which achieves adaptive aggregation in space,thereby enhancing the model’s adaptability to geometric transformations and feature representation capabilities. Secondly,by introducing SimAM attention mechanism,the feature extraction ability of the backbone is further strengthened,improving the sensitivity and recognition ability of the network to input data,enabling the model to support more complex visual tasks. The experimental results show that the accuracy of the TODC-YOLOv9 model reaches 98. 02% ,the recall rate reaches 93. 75% ,and the average accuracy mean mAP_0. 5 and mAP_0. 5:0. 95 reach 98. 27% and 88. 72% ,respectively,which are better than that of other models. It is indicated that the TODC-YOLOv9 model has excellent robustness and recognition ability,which can provide strong technical support for tomato disease detection in complex back-grounds.