基于非局部支持注意力的小样本目标检测算法  

Non-Local Support Attention Network for Few-Shot Object Detection

作　　者：谢熙君 李菲菲[1] XIE Xijun;LI Feifei(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《电子科技》2024年第8期75-83,共9页Electronic Science and Technology

基　　金：上海市高校特聘教授(东方学者)岗位计划(ES2015XX)。

摘　　要：基于元学习的小样本目标检测算法研究的关键之处,是更好地利用支持分支的信息来更有效地帮助查询分支完成对新类目标的识别,较多算法在查询分支加入支持分支信息时只在深度方向进行融合,忽略了特征之间的空间位置关系。文中提出基于非局部支持注意力的小样本目标检测算法模型,该方法不仅在候选框特征中加入了支持信息,还将支持信息与送入候选框生成网络的特征进行融合,同时考虑了特征之间的空间位置关系,在检测模块中加入负支持样本的信息帮助模型区分异类目标。该模型在COCO2017数据集的基类和新类上均获得了良好的检测效果。在增量式学习的情况下,相比改进前,在新类AP(Average Precision)/AP_(50)/AP_(75)上分别增加了3.3/3.8/4.7 mAP(mean Average Precision),在基类AP/AP_(50)/AP_(75)上分别增加了2.7/0.5/3.3 mAP,并且超过了相同设置下SOTA(Sort-Of-The-Art)模型DAnA(Dual-Awareness Attention)的表现。The key point of current research on few-shot object detection based on meta-learning is how to make better use of the information of support branch to help query branch to recognize novel objects more effectively.However,many current methods fuse the features from support branch and query branch in the depth direction,ignoring the spatial position relationship between features.Therefore,this study proposes non-local support attention network.This method not only adds support information into the proposals features,but also fuses the support information with the features that fed into region proposal network.The spatial position relationship between features is considered at the same time.It also adds the information of negative supports to the detection module to help the model distinguish the objects from different categories.This method obtains good performance on base classes and novel classes of COCO2017 dataset,particularly under the case of incremental learning.Compared with method before improvement,3.3/3.8/4.7 mAP is increased in AP/AP_(50)/AP_(75)of novel classes.2.7/0.5/3.3 mAP is increased in AP/AP_(50)/AP_(75)of the base classes,and outperformed the performance of SOTA(Sort-Of-The-Art)model DAnA(Dual-Awareness Attention)under the same setting.

关 键 词：目标检测 小样本学习 元学习 增量式学习 特征融合 注意力 非局部 微调 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]