【作者】 康彬； 赵世豪； 李鑫； 汪进；

【Author】 Bin Kang;Shihao Zhao;Xin Li;Jin Wang;School of Internet of Things, Nanjing University of Posts and Telecommunications;Jiangsu Academy of Chemical Inherent Safety;

【机构】 南京邮电大学物联网学院； 江苏省化工本质安全研究院；

【摘要】 细粒度识别关注目标间的细节差异,在样本标记时需要专业指导。为了降低训练成本,基于小样本的细粒度识别已成为研究热点。本文提出了一个粗标签引导的小样本细粒度识别方法。所提方法旨在有效利用易获取的粗粒度标记样本预训练网络,以此弥合小样本学习与全监督学习之间巨大的性能差异。具体而言,论文设计了粗-细粒度标签交替学习的网络架构,所提网络架构能有效利用粗-细粒度样本特征图的相似性降低识别网络对细粒度样本数量的依赖。所设计网络主要包括三个组件:1)基于粗粒度标签预训练的卷积神经网络架构;2)预测粗-细粒度特征图一致性的网络模块;3)提升模型泛化能力的掩模生成方法。除此之外,所设计网络还通过设计队列进出站策略解决粗-细标记类别不平衡问题。通过三个数据集的大量实验表明在只保留20%细标签的情况下,论文所提方法相较于主流全监督细粒度识别网络,识别精度仅降低6%。更多还原

【Abstract】 Fine grained recognition focuses on the detailed differences between targets, so professional guidance is required when labeling samples. In order to reduce training costs, fine-grained recognition based on small samples has become a research hotspot. This article proposes a coarse label guided fine-grained recognition method for small samples. The proposed method aims to effectively use the easily available coarse grained labeled sample pre training network to bridge the huge performance difference between small sample learning and full Supervised learning. Specifically, the paper designs a network architecture for alternating learning of coarse-grained and fine-grained labels. The proposed network architecture can effectively utilize the similarity of coarse-grained sample feature maps to reduce the dependence of the recognition network on the number of fine-grained samples. The designed network mainly includes three components: 1) Convolutional neural network architecture based on coarse grained label pre training; 2) A network module for predicting the consistency of coarse-grained feature maps; 3) A mask generation method to enhance model generalization ability. In addition, the designed network also solves the problem of imbalanced coarse fine labeling categories by designing queue inbound and outbound strategies. Extensive experiments on three datasets have shown that the proposed method reduces recognition accuracy by only 6% compared to mainstream fully supervised fine-grained recognition networks, while retaining only 20% fine labels.更多还原