基于改进EfficientNet的辣椒病虫害识别方法研究  

Research of the pepper diseases and pests recognition methods based on improved EfficientNet

作　　者：张振强 赵伟峰[1] ZHANG Zhenqiang;ZHAO Weifeng(School of Management,Anhui Science and Technology University,Bengbu Anhui 233100,China)

机构地区：[1]安徽科技学院管理学院,安徽蚌埠233100

出　　处：《延边大学农学学报》2025年第1期46-50,共5页Journal of Agricultural Science Yanbian University

基　　金：安徽省重大项目(2022AH040230);安徽科技学院校级科研发展基金项目(FZ230122)。

摘　　要：该研究针对辣椒病虫害识别过程中数据量少、精度低等问题,以安徽科技学院农学院辣椒试验田采集的病虫害图片数据为研究对象,利用多尺度注意力机制来改进EfficientNet模型,实现了辣椒病虫害的高效精确识别。改进后的模型不仅可以增强和突出病虫害区域,而且可以从多个角度提取图像特征,从而提高了模型对特征的捕获能力。结果表明:改进后的EfficientNet识别准确率达到了94.13%,与原模型相比,识别准确率提升了3.60个百分点。与流行的VGG-19、ResNet-50和DenseNet-121模型相比,准确率分别高出了28.63、9.54和8.20个百分点,证明了该方法的有效性。This study aims to address the issues of limited data volume and low accuracy in the traditional pepper diseases and pests recognition.Taking the diseases and pests image data collected from the pepper experimental field of the College of Agriculture,Anhui Science and Technology University as the research object,a multi-scale attention mechanism is used to improve the EfficientNet model,achieving efficient and accurate recognition.The improved model can not only enhance and highlight pest and disease areas,but also extract image features from multiple perspectives,thereby improving the ability to capture features.The experimental results show that the improved EfficientNet has an accuracy rate of 94.13%,which is 3.60%higher than the original model in terms of recognition accuracy.Compared with the popular VGG-19,ResNet-50,and DenseNet-121 models,the accuracy rates increased by 28.63%,9.54%,and 8.20%,respectively,demonstrating its effectiveness.

关 键 词：辣椒病虫害识别 多尺度注意力机制 EfficientNet 深度学习 

分 类 号：S436.418[农业科学—农业昆虫与害虫防治]