基于多尺度特征拼接的小样本茶叶病害分类  

Few-shot tea disease classification based on multi-scale feature splicing

作　　者：张艳[1] 王林茂 程志友[1] 章杨凡 储著增 ZHANG Yan;WANG Linmao;CHENG Zhiyou;ZHANG Yangfan;CHU Zhuzeng(College of Electronics and Information Engineering,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学电子信息工程学院,安徽合肥230601

出　　处：《安徽大学学报（自然科学版）》2022年第5期58-63,共6页Journal of Anhui University(Natural Science Edition)

基　　金：国家重点研发计划重点专项(2018YFC0807302);国家自然科学基金资助项目(61772032);基本科研业务费项目(2018JB08);安徽省高等学校自然科学研究项目(KJ09A0027)。

摘　　要：传统的茶叶病害分类是一项耗时耗力的工作.针对该问题,提出一种基于多尺度特征拼接的网络模型,用于小样本茶叶病害分类.通过多尺度注意力模块提取茶叶叶片的显著性特征,进而得到显著性图像.对显著性图像与原始图像进行通道特征拼接,使拼接后的图像既包含全局特征又包含局部特征.融合多个不同卷积层输出的特征,使特征图包含空间和语义信息.分类实验结果表明:用可分离卷积代替常规卷积后,该文模型参量总数小于关系网络模型参量总数的1/2,提高了分类效率;相对于其他5种模型,该文模型分类准确率最高.The traditional classification of tea diseases is a time-consuming and labor-intensive task.A network model was proposed in order to solve this problem,which was based on multi-scale feature splicing and was used for classification of small samples of tea diseases.The saliency features of tea leaves were extracted through the multi-scale attention module,and then saliency images were obtained.The saliency image and the original image were stitched with channel features,so that the image which was stitched could contain both global features and local features.The output features by multiple different convolutional layers were fused,so that the feature map could contain spatial and semantic information.The results of classification experiments showed that after replacing conventional convolution with separable convolution,the total number of parameters of this paper model was less than half of the total number of parameters of the relational network model,which could improve the classification efficiency.Compared with other five models,the model of this paper had the highest classification accuracy.

关 键 词：茶叶病害分类 多尺度注意力模块 显著性区域 可分离卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]