基于视觉一致性增强的细粒度图像检索  

Fine-grained Image Retrieval Based on Strengthened Visual Consistency

作　　者：郎文溪 孙涵[1] LANG Wen-xi;SUN Han(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2022年第12期12-20,共9页Computer Technology and Development

基　　金：国防科技创新特区项目资助(XX);中央高校基本科研业务费专项资金(NZ2019009)。

摘　　要：细粒度图像检索旨在从大类图像中检索出特定子类的图像。得益于卷积神经网络的快速发展,细粒度图像检索的精度和速度均取得突破,但其性能仍受限于不同子类图像间高相似性和同一子类图像间的高差异性。针对上述问题,该文提出了一种基于对比学习和视觉一致性增强的细粒度图像检索框架CVCS-Net。CVCS-Net由判别性特征挖掘模块、视觉一致性增强模块和语义哈希编码模块组成,在挖掘类间图像判别性特征的同时,通过增强类内图像的视觉一致性来提升模型对类内图像差异的容忍度。判别性特征挖掘模块学习空间注意力图来定位图像的判别性区域并获得这些区域对应的局部特征表示;视觉一致性增强模块提升模型对类内图像差异的鲁棒性;而语义哈希编码模块基于量化损失和位平衡损失进一步学习紧凑的哈希码用于检索。CVCS-Net在CUB200-2011、Stanford Dogs和Stanford Cars的mAP分别可达到0.8591、0.8564和0.9183,相较于当前其他检索方法能够取得更好的检索结果。Fine-grained image retrieval aims at retrieving images of specific sub-categories from general categories of images.Thanks to the rapid development of convolutional neural networks,there has been a breakthrough in the accuracy and speed of fine-grained image retrieval.However,its performance is still limited by the high similarity between images of different sub-categories and the high difference between images of the same sub-category.Therefore,a contrast learning and strengthened visual consistency CVCS-Net is proposed.CVCS-Net consists of three key modules:discriminative feature mining,strengthened visual consistency and semantic hash coding.The discriminative feature mining module learns spatial attention maps to locate discriminative regions of images and obtains local feature representations corresponding to these regions;the strengthened visual consistency module improves the robustness of the model to intra-class image differences;and the semantic hash coding module further learns compact hash codes for retrieval based on quantization loss and bit balance loss.CVCS-Net can get mAPs of 0.8591,0.8564 and 0.9183 for CUB200-2011,Stanford Dogs and Stanford Cars,respectively,which can get better results compared with other current retrieval methods.

关 键 词：细粒度图像检索 弱监督 对比学习 哈希 视觉一致性 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]