摘要
针对玉米叶部病害之间差异较小,难以快速、准确区分的问题,提出了一种高性能的轻量化玉米病害识别模型CB-SV-MobileNetV3,通过在原MobileNetV3-Small网络结构的基础上进行调整。首先,将原模型中的ReLU激活函数替换为ELU激活函数,改善模型的非线性表达能力和缓解梯度消失问题。其次,在Bottleneck结构中,使用CBAM注意力机制替换SE注意力机制,提升模型对玉米病害特征区域的关注力和表达能力,并添加ResMLP结构,加强模型对玉米病害特征的捕获能力。最后,将SVM作为玉米病害的分类器,进一步提高模型对玉米病害的分类精度。结果表明,CB-SV-MobileNetV3模型在公开玉米叶部病害数据集上识别准确率为97.49%,优于AlexNet、MobileNetV2等模型。在自建玉米叶部病害数据集上准确率达到91.23%,且该模型参数量仅为2.96 MB,检测帧率为69 frame/s。
In view of the problem that it is difficult to quickly and accurately distinguish between different types of maize leaf diseases due to the small differences between them,this study proposes a high-performance lightweight maize disease recognition model CB-SV-MobileNetV3.By adjusting the original MobileNetV3-Small network structure,the ReLU activation function in the original model is replaced with the ELU activation function to improve the model's nonlinear expression ability and alleviate the problem of vanishing gradients.Secondly,in the Bottleneck structure,the CBAM attention mechanism is used instead of the SE attention mechanism to enhance the model's attention and expression ability for maize disease feature regions,and the ResMLP structure is added to strengthen the model's ability to capture maize disease features.Finally,SVM is used as a classifier for maize diseases to further improve the model's classification accuracy for maize diseases.The results show that the CB-SV-MobileNetV3 model has an accuracy rate of 97.49%on the publicly available maize leaf disease dataset,which is superior to models such as AlexNet and MobileNetV2.On the selfbuilt maize leaf disease dataset,the accuracy rate reaches 91.23%,and the model has a parameter count of only 2.96 MB,with a detection frame rate of 69 frame/s.
作者
曾鹏滔
撒金海
余兰灿
刘嘉
ZENG Pengtao;SA Jinhai;YU Lancan;LIU Jia(College of Software,Xinjiang University,Urumqi 830008,China)
出处
《微电子学与计算机》
2025年第4期98-105,共8页
Microelectronics & Computer
基金
国家自然科学基金(62266043)。
