可变形自注意力的自适应原型算子互学习模型  

Adaptive prototype operator mutual learning model with deformable self-attention

作　　者：吕佳[1,2] 董保森 曾梦瑶 LU Jia;DONG Baosen;ZENG Mengyao(College of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China;Chongqing Engineering Technology Research Center of Digital Agriculture Services,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331 [2]重庆师范大学重庆市数字农业服务工程技术研究中心,重庆401331

出　　处：《武汉大学学报(工学版)》2025年第3期470-481,共12页Engineering Journal of Wuhan University

基　　金：国家自然科学基金重大项目(编号:11991024);重庆市教委“成渝地区双城经济圈建设”科技创新项目(编号:KJCX2020024);重庆市教委科研项目重点项目(编号:KJZD-K202200511);重庆市高校创新研究群体资助项目(编号:CXQT20015)。

摘　　要：针对基于度量学习的小样本学习分类模型存在对样本关键特征提取能力不足以及原型在度量空间中特征姿态完整性缺失的问题,提出了一个基于可变形自注意力的自适应原型算子互学习模型。利用可变形卷积自适应地将自注意力中查询向量和键值向量的感受野引导至重点区域,为自注意力提供更准确的上下文语义信息,进而动态捕获样本的关键特征;在考虑原型全局信息的基础上构建多个局部原型算子以获取局部信息,利用互学习的思想融合原型特征的全局信息和局部信息,补全原型在度量空间中的特征姿态;联合全局损失函数和局部损失函数对模型参数进行反向更新。文中使用ResNet10和ResNet12分别在2个小样本公用数据集上进行实验:在miniImageNet数据集上,所提出模型的分类准确率相较基准提升了2.31%~2.58%;在CUB数据集上,所提出模型的分类准确率相较基准提升了1.38%~2.96%。实验表明,该模型在不同粒度的数据集上,针对不同的骨干网络均有不同程度的性能提升。Aiming at the issues that the few-shot learning classification models based on metric learning lack the ability to extract key features from samples and the prototypes lack the integrity of feature pose in the metric space,an adaptive prototype operator mutual learning model based on deformable self-attention is proposed.Firstly,the receptive fields of query vector and key-value vector in self-attention are adaptively guided to the key areas by using deformable convolution,so as to provide more accurate contextual semantic information for selfattention,and then dynamically capture the key features of samples.Secondly,on the basis of considering the global information of the prototype,multiple local prototype operators are constructed to obtain the local information,and the idea of mutual learning is utilized to fuse the global information and local information of the prototype feature to complete the feature pose of the prototype in the metric space.Finally,the global loss function and the local loss function are combined to reversely update the model parameters.In this paper,ResNet10 and ResNet12 are used to conduct experiments on two few-shot public datasets,respectively.On the mini-ImageNet dataset,the classification accuracy is improved by 2.31%to 2.58%compared with the baseline.On the CUB dataset,the classification accuracy is improved by 1.38%to 2.96%compared with the baseline.Experiments show that the presented model has different degrees of performance improvement for different backbone networks on datasets with different granularities.

关 键 词：小样本学习 度量学习 元学习 可变形自注意力 互学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]