用于植物病虫害图像识别的数据增强方法  

A Data Augmentation Method for Image Recognition of Plant Pests and Diseases

作　　者：肖宇[1] 吴杰[1] 马驰 XIAO Yu;WU Jie;MA Chi(School of Computer and Software Engineering,University of Science and Technology Liaoning,Anshan 114051,China;School of Computer Scienceand Engineering,Huizhou University,Huizhou 516007,China)

机构地区：[1]辽宁科技大学计算机与软件工程学院,辽宁鞍山114051 [2]惠州学院计算机科学与工程学院,广东惠州516007

出　　处：《计算机技术与发展》2025年第3期210-214,共5页Computer Technology and Development

基　　金：广东省教育科学研究项目(2022ZDZX4052,2021ZDJS082)。

摘　　要：在深度学习的植物病虫害图像识别领域,区域数据增强是提高模型泛化能力的关键策略。该技术通过有选择性地移除图像的特定区域,促使模型更加专注于提取那些区分度较低的特征,进而增强了模型对新数据的适应性和识别能力。所提出的SaliencyBatchMix数据增强方法利用类激活映射(CAM)计算语义百分比图(SPM)后,按Batch的维度选择具有代表性的图像区域,并将图像区域作为指示性补丁与目标图像混合,以引导模型学习更恰当的特征表示。该方法可减少训练中裁剪区域的无意义像素,并且减少了标签噪声。在GhostNet架构下的实验中,SaliencyBatchMix分别在IP102和Embrapa数据集上实现了72.05%和96.86%的准确率。对比于使用CutMix方法分别提升了0.62百分点和1百分点。通过对结果的对比和消融实验分析,验证了SaliencyBatchMix在提高模型泛化能力和准确率的有效性。In the field of deep learning image recognition of plant pests and diseases,regional data augmentation is a key strategy to improve model generalization.It prompts models to focus on extracting less discriminative features by selectively removing image regions,thereby improving adaptability to new data.The proposed SaliencyBatchMix method utilizes Class Activation Mapping(CAM)to derive Semantic Percentage Maps(SPM)after selecting representative regions based on batch dimensionality.It mixes indicative patches into target images to guide learning more suitable representations.This reduces meaningless cropped pixels during training and labeling noise.Experiments using GhostNet show SaliencyBatchMix achieves 72.05%and 96.86%accuracy on IP102 and Embrapa,outperforming CutMix by 0.62 percentage points and 1 percentage points,respectively.Results validation and ablation study findings corroborate SaliencyBatchMix's effectiveness in boosting model generalization and accuracy.

关 键 词：数据增强 类激活映射 深度学习 植物病虫害识别 GhostNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]