面向细粒度图像的数据关联代理损失

Data relation proxy loss for fine-grained image

现有度量学习方法中基于元组的损失训练速度慢、基于代理的损失未考虑数据间细粒度的语义关系。针对这些问题,结合两者的优势提出了一种面向细粒度图像的数据关联代理损失(data relation proxy loss,DRPLoss)函数。采用具有批量归一化(BN)层的inception网络作为嵌入网络,在度量空间中利用梯度相互交互学习数据间的相关关系,并使用温度缩放调节DRPLoss对嵌入向量进行监督训练。在CUB-200-2011和Car-196数据集上验证了不同嵌入维度的DRPLoss的有效性,recall@1评价指标分别提升了2%和6.4%。实验结果表明,相比基于元组的损失和基于代理的损失,DRPLoss的训练速度更快,对细粒度图像检索的性能有显著性提高。

In the existing metric learning methods,the tuple-based loss training speed is slow and the proxy-based loss don’t consider the fine-grained semantic relationship between the data.In response to these problems,the paper combined the advantages of the two and proposed a DRPLoss function for fine-grained images.This paper used an inception network with a BN layer as the embedding network,learnt the correlation between data through gradient interaction in the metric space,and used temperature scaling to adjust the DRPLoss to supervise and train the embedding vector.Finally,this paper verified the effectiveness of diffe-rent embedding dimensions DRPLoss on CUB-200-2011 and Car-196 datasets.The experiment improves the evaluation index of recall@1 by 2%and 6.4%respectively.Compared with the tuple-based loss and proxy-based loss,the experimental results show that DRPLoss is faster in training and has a significant improvement in the performance of fine-grained image retrieval.