基于元学习的植物虫害识别原型网络VGG-ML

Plant pest identification prototype network VGG-ML based on meta-learning

[目的]为解决使用深度学习技术对植物虫害识别时依赖大量训练样本的问题,本文基于元学习的思想设计一个VGG原型网络(VGG-meta learning,VGG-ML),用于在小样本背景下植物虫害种类识别。[方法]采用VGG16作为嵌入单元提取虫害样本特征与类别特征,为提高网络对于新类别的识别能力,采用训练集与测试集异域方式进行模型训练,以解决在小样本情况下植物虫害识别准确率低、新类别虫害无法识别的问题。将测试集划分为支持集(获取类原型)与查询集(样本原型),以欧式距离度量样本原型与类原型之间的相似性,从而判定样本所属类别。[结果]以公开数据集IP102中玉米、甜菜、苜蓿等11种植物的蚜虫、黏虫、跳甲等24类农业虫害图片作为训练数据,以稻纵卷叶螟、稻叶毛虫、亚洲稻螟、稻瘿蚊、稻秆蝇、稻水象甲、稻叶蝉、稻苞虫8类常见的水稻虫害作为测试数据,在5-way、1-shot与5-way、5-shot情况下VGG-ML识别准确率分别为67.98%与81.5%,与原始原型网络相比提高3.53与4.4百分点。5-way、5-shot试验与基于迁移学习的ResNet50与VGG16网络对比,准确率分别提高28.65与25.94百分点。[结论]VGG-ML在进行小样本植物虫害类型识别时有效可靠,可适用于小样本植物识别问题。

[Objectives]To solve the problem of relying on a large number of training samples when using deep learning technology to identify plant pests,a VGG(visual geometry group)prototype network(VGG-meta learning,VGG-ML)based on the idea of meta-learning was proposed in this pater to identify plant pest types in small sample backgrounds.[Methods]VGG16 was used as the embedding unit to extract the characteristics and category characteristics of the plant pest sample.In order to improve the recognition ability of the network for new categories,and solve the problem of low recognition accuracy of plant pests and unrecognizable new categories of pests in the case of small samples,the dataset that the training set and the test set from different data categories was adopted in this pater.The test set was divided into a support set(obtaining class prototypes)and a query set(sample prototypes),and the similarity between sample prototypes and class prototypes was measured by Euclidean distance to determine the category to which the samples belong.[Results]Twenty four kinds of agricultural insect pests such as aphids,armyworms and flea beetles of 11 plants such as corn,sugar beet,and alfalfa in the public dataset IP102 were used as training data,and 8 kinds of common aquatic rice pests such as rice leaf roller,rice leaf caterpillar,Asian rice borer,rice gall midge,rice stem fly,rice water weevil,rice leaf hopper,and rice bract were used as test data.The recognition accuracy of VGG-ML was 67.98% and 81.5% respectively under 5-way,1-shot and 5-way,5-shot conditions,which was 3.53 and 4.4 percentage points higher than the original prototype network,respectively.Compared with the ResNet50 and VGG16 networks based on transfer learning,the accuracy of the 5-way and 5-shot tests increased by 28.65 and 25.94 percentage points,respectively.[Conclusions]VGG-ML was effective and reliable in the identification of plant pest types in small samples,and it could be applied to the identification of small samples of plants.