河南农业大学学报2024，Vol.58Issue(6)：1021-1033,13.DOI:10.16445/j.cnki.1000-2340.20240429.003
基于多尺度特征融合和注意力机制的辣椒病害识别模型
Pepper disease identification model based on multi-scale feature fusion and attention mechanism
摘要
Abstract
[Objective]The MobileNet with large convolution Unit(Mobile-LU)model was designed to solve the problems of difficulty in disease identification and low accuracy due to the complexity of dis-ease types and the lack of obvious differences between classes.[Method]The feature extraction layer of MobileNetV3 was restructured to enhance the model's ability to express the features of varying sizes in pepper diseases by employing separable convolutions of different scales in the parallel branch units.The Squeeze-and-Excitation(SE)attention mechanism was introduced to strengthen the model's learning of disease-related features and improve the accuracy of disease identification.Additionally,the Leaky ReLU activation function was employed with a small slope introduced in the negative region to prevent neuron death in the network.The node count in the output layer was adjusted to better adapt to pepper disease classification tasks.[Result]The Mobile-LU model achieves a recognition accuracy of 98.2%,surpassing the MobilenetV3-small,ResNet34,VGG16,Alexnet,Swin Transformer,and MobileVIT mod-els by 8.9,7.3,4.4,20.4,6.0,8.3 percentage points,respectively.Moreover,the Mobile-LU model demonstrates advantages in key performance indicators such as precision,recall,specificity,and F1 score.[Conclusion]The Mobile-LU model exhibits superior performance in pepper disease recognition,and can better meet the requirements of pepper disease identification tasks.关键词
辣椒病害/图像分类/SE注意力机制/深度可分离卷积/多尺度特征融合Key words
pepper diseases/image classification/SE attention mechanism/depth separable convolu-tion/multi-scale feature fusion分类
信息技术与安全科学引用本文复制引用
尚俊平,张冬阳,席磊,刘合兵,苏楠..基于多尺度特征融合和注意力机制的辣椒病害识别模型[J].河南农业大学学报,2024,58(6):1021-1033,13.基金项目
河南省研究生教育改革与质量提升工程项目(YJS2023AL046) （YJS2023AL046）
河南省现代农业产业技术体系(S2010-01-G04) （S2010-01-G04）
