基于空间注意力增强ResNeSt-101网络和迁移元学习的小样本害虫分类  

Few-shot pest classification using spatial-attention-enhanced ResNeSt-101 network and transfer-based meta-learning

作　　者：梁炜健 郭庆文 王春桃[1,2,3,4] 肖德琴 黄琼[1,2,4] LIANG Weijian;GUO Qingwen;WANG Chuntao;XIAO Deqin;HUANG Qiong(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China;Key Laboratory of Smart Agricultural Technology in Tropical South China,Ministry of Agriculture and Rural Affairs,Guangzhou 510642,China;Guangdong Provincial Key Laboratory of Agricultural Artificial Intelligence,Guangzhou 510642,China;Guangzhou Key Laboratory of Intelligent Agriculture,Guangzhou 510642,China;Shenzhen Pengcheng Technician College,Shenzhen 518000,China)

机构地区：[1]华南农业大学数学与信息学院,广州510642 [2]农业农村部华南热带智慧农业技术重点实验室,广州510642 [3]广东省农业人工智能重点实验室,广州510642 [4]广州市智慧农业重点实验室,广州510642 [5]深圳鹏城技师学院,深圳518000

出　　处：《农业工程学报》2024年第6期285-297,共13页Transactions of the Chinese Society of Agricultural Engineering

基　　金：National Science Foundation of China(62172165);Science and Technology Planning Project of Guangdong Province under Grant(2021B1212040009);Natural Science Foundation of Guangdong Province(2022A1515010325);Guangzhou Basic and Applied Basic Research Project(202201010742);Science and Technology Program of Guangzhou(202206010116,201902010081,107126242281)。

摘　　要：害虫识别是害虫防治的关键基础,由于较难获得足够的害虫种类图像,如何使用少量标记图像构造害虫分类器是一个富有挑战性的问题。现有研究多采用匹配网络框架来解决这个问题,该框架使用元学习避免重新训练深度网络,然而主干网络的特征提取能力有限,元学习算法没有提供较好的权重初始化策略,可能导致网络出现梯度消失或者梯度爆炸的情况。为了解决这一问题,该研究提出一种基于空间注意力增强ResNeSt-101和迁移元学习算法的小样本害虫分类器。首先,通过一个空间注意力模块增强ResNeSt-101以更好地提取害虫图像特征,即在ResNeSt-101的第1阶段的最大池化层之前以及在第2~4阶段的末尾分别附加集成空间注意力模块,并通过数值仿真确定空间注意力增强模块的最佳放置位置为第1阶段的最大池化层之前。随后,通过迁移学习策略初始化网络权重,进而通过元学习进行优化。为了避免网络出现梯度消失或者梯度爆炸的情况,在元学习算法中选择归一化的温度缩放交叉熵损失函数代替三元组损失函数。最后,通过计算查询图像和支持图像深度特征之间的相似度实现害虫分类。所提出方法在自建的害虫图像数据集AD0和MIP50上使用N-类K-例准确率和每张图像处理时间(the time of per image processing,TPIP)进行评估。害虫图像数据集的构建方式如下:首先对公共害虫图像数据集IP102和D0进行清洗,以消除由于英文害虫名称导致的歧义类别;然后移除卵、幼虫和蛹阶段的害虫图像,仅保留成虫阶段的图像。考虑到人工和时间成本,从清理后的IP102害虫数据集中选择50个类别构建MIP50害虫图像数据集。随后,通过害虫的拉丁名称从互联网搜索更多的害虫图像,生成AD0害虫图像数据集。自建的MIP50数据集包括来自IP102的50个类别的16424张成虫图像,AD0包含来自D0的所有40个类别的17112�Pest recognition is a key foundation of pest management.Previous researches have exploited image classification to achieve automatic pest recognition.As it is difficult to obtain sufficient images for new-emerged pest classes,how to develop a pest classifier with a few labeled images is an interesting and challenging problem.Some existing works in the literature employ the matching network framework to solve this problem,which use metalearning to avoid retraining deep networks.In these works,however,the feature extraction abilities of backbone networks are limited and the meta-learning algorithms do not provide a good weight initialization strategy or might cause network collapse.To close this gap,a few-shot pest classifier using a spatial-attention-enhanced version of ResNeSt-101 and a transfer-based meta-learning algorithm is proposed in this study.First,ResNeSt-101 is enhanced with a spatial attention block to better extract image features.The spatial attention block is suggested to integrate before the max pooling layer in the first stage of ResNeSt-101 and/or append at the end of stages 2-4,and the optimal location is determined as the first stage via the numerical simulation results.Subsequently,network weights are initialized by transfer learning and then optimized by meta-learning.To avoid network collapse,the normalized temperature-scaled cross-entropy loss function instead of the triplet loss function is chosen in the meta-learning algorithm.Finally,pest classification is achieved by computing similarities between deep features of query and support images.In addition,the proposed method is evaluated on two elaborately constructed pest image datasets AD0 and MIP50 with N-way K-shot accuracy and the time of per image processing(TPIP).These two pest image datasets are constructed as follows:images in the public pest image datasets,IP102 and D0,are firstly cleaned by eliminating the images with class ambiguities due to the English pest name-based categorizing;and the images of eggs,larvae,and pupae stages a

关 键 词：病虫害 图像处理 小样本分类 元学习 ResNeSt-101 交叉熵损失 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S24[自动化与计算机技术—计算机科学与技术]