基于联合优化多任务学习的细粒度图像识别  

Fine-grained image recognition based on joint optimization multi-task learning

作　　者：朱阳光 刘瑞敏[1] 王震 王枭 ZHU Yang-guang;LIU Rui-min;WANG Zhen;WANG Xiao(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院

出　　处：《陕西理工大学学报（自然科学版）》2019年第6期41-47,共7页Journal of Shaanxi University of Technology:Natural Science Edition

摘　　要：为了能够在细粒度图像特征表示中探索出相似层结构中的共享信息,提出了一种多任务学习框架,联合优化卷积神经网络中的Softmax Loss和Triplet Loss,基于此框架,设计广义三元组嵌入标签结构,以发现不同级别中具有相似性的相关图像。在Stanford Cars和CUB200-2011两个细粒度数据集上进行实验,结果表明这种方法不仅可以实现较好的分类性能,还能够提高在细粒度数据集上不同级别的标签结构的图像检索精度,这在电子商务中相关产品的推荐方面具有重要意义。In order to explore the shared information in the similar layer structure in the fine-grained image feature representation, a multi-task learning framework is proposed to jointly optimize Softmax Loss and Triplet Loss in the convolutional neural network. Based on this framework, a generalized triple is designed to embed the label structure to find related images with similarities at different levels. Experimenting on two publicly fine-grained data sets, Stanford Cars and CUB200-2011 shows that this method can not only achieve better classification performance, but also improve the image retrieval precision of different levels of label structure on fine-grained data sets. This will have great significance in the recommendation of related products in e-commerce.

关 键 词：细粒度图像分类 深度神经网络 图像识别 联合优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]