基于FREW-DT的玉米叶片病害检测
集成电路应用
卢志强,孙俊,彭逸飞,高锦烨
江苏大学电气信息工程学院
摘要: 为解决复杂背景下玉米叶片病害目标小、形状不一、易受背景干扰等目标检测难题,提出了一种基于改进RT-DETR的玉米叶片病害深度学习模型FREW-DT。该研究对RT-DETR模型在三个方面进行改进:采用FasterNet block重构主干网络以减少计算冗余;通过改进的RepVGG block优化颈部网络提升推理速度;在特征融合模块引入EMA注意力机制增强跨尺度特征交互。实验结果表明,FREW-DT模型参数量为31.09 M,平均精度均值达95.36%,精确率和召回率分别为90.85%和86.72%,帧率提升了29 FPS。该模型可高效实现玉米叶片病害精准检测,为现代农业病害防治提供技术支持。
中图分类号:TP391.41文献标识码:ADOI:10.19339/j.issn.1674-2583.2026.03.010
中文引用格式:卢志强,孙俊,彭逸飞,等. 基于FREW-DT的玉米叶片病害检测[J].集成电路应用,2026,43(3):48-53.
英文引用格式:Lu Zhiqiang, Sun Jun, Peng Yifei, et al. Maize leaf disease detection model based on FREW-DT[J].Application of IC,2026,43(3):48-53.
中文引用格式:卢志强,孙俊,彭逸飞,等. 基于FREW-DT的玉米叶片病害检测[J].集成电路应用,2026,43(3):48-53.
英文引用格式:Lu Zhiqiang, Sun Jun, Peng Yifei, et al. Maize leaf disease detection model based on FREW-DT[J].Application of IC,2026,43(3):48-53.
Maize leaf disease detection model based on FREW-DT
Lu Zhiqiang, Sun Jun, Peng Yifei, Gao Jingye
School of Electrical and Information Engineering, Jiangsu University
Abstract: To address the challenges of maize leaf disease detection, such as small target size, irregular shapes, and susceptibility to background interference in complex environments, this paper proposes a deep learning model FREW-DT for maize leaf disease detection based on the improved RT-DETR. Three improvements are made to the RT-DETR model in this study: reconstructing the backbone network with FasterNet blocks to reduce computational redundancy, optimizing the neck network with improved RepVGG blocks to enhance inference speed, and introducing the EMA attention mechanism into the feature fusion module to strengthen cross-scale feature interaction. The experimental results show that the FREW-DT model has a parameter count of 31.09 M, a mean average precision (mAP) of 95.36%, a precision of 90.85%, and a recall of 86.72%, with the frame rate increased by 29 FPS. This model can efficiently achieve accurate detection of maize leaf diseases and provide technical support for disease prevention and control in modern agriculture.
Key words : object detection;maize leaf diseases; RT-DETR; deep learning
引言
玉米作为国家粮食安全的战略农作物,扮演重要角色。然而气候变暖导致病害频发,科技赋能现代农业体系日益重要。
最初国内外研究人员主要依靠传统的机器学习[1]算法识别、检测目标[2]。随着计算机技术的发展,深度学习算法的应用逐渐广泛[3]。孙俊团队[4]在MobileNetV2中构建了协调注意力机制,通过改进位置敏感特征并结合多尺度特征,在复杂背景下农作物病害识别的准确率达到92.2%,模型参数数量减少了76%。Liu等人[5]通过迁移学习策略将MobileNetV2嵌入YOLOv3,在番茄灰腐病的识别中精度达到94.3%。
现有主流模型在复杂农田背景下,对小尺寸、形态不规则的玉米病斑存在漏检率高、精度不足、速度与轻量化难以兼顾等问题。玉米叶片病害在叶片中所占面积较小,背景复杂,为了实现复杂背景下玉米叶片病害便携、精准地检测,本研究提出一种改进的RTDETR[6]玉米叶片病害检测模型,通过改进主干网络、颈部网络并引入注意力机制,提高模型的检测精度,为现代农业高质量发展助力。
本文详细内容请下载:
https://www.chinaaet.com/resource/share/2000007198
投稿邮箱:ICAPP@belling.com.cn
作者信息:
卢志强,孙俊,彭逸飞,高锦烨
(江苏大学电气信息工程学院,江苏镇江212013)

此内容为AET网站原创,未经授权禁止转载。
