基于偏移过滤与未知特征强化的开放世界目标检测  

Offset filter and unknown feature reinforcement for open world object detection

作　　者：王超 苏树智 朱彦敏 徐阳 Wang Chao;Su Shuzhi;Zhu Yanmin;Xu Yang(School of Computer Science&Engineering,Anhui University of Science&Technology,Huainan Anhui 232001,China;School of Mechanical&Electrical Engineering,Anhui University of Science&Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]安徽理工大学机电工程学院,安徽淮南232001

出　　处：《计算机应用研究》2025年第2期618-622,共5页Application Research of Computers

基　　金：国家自然科学基金面上项目(52374155,61806006);安徽省自然科学基金(面上项目)(2308085MF218);安徽省高等学校自然科学研究项目(重大项目)(2022AH040113);安徽理工大学青年基金(重点项目)(QNZD202202);淮南市指导性科技计划项目(2023142,2023147);安徽省高校中青年教师培养行动项目(YQZD2023035);安徽理工大学医学专项培育项目(重大项目)(YZ2023H2A007);合肥综合性国家科学中心大健康研究院职业医学与健康联合研究中心项目(OMH-2023-05,OMH-2023-24)。

摘　　要：开放世界目标检测(open world object detection,OWOD)是一个计算机视觉挑战,聚焦于现实世界环境,其不仅要检测出标记出的已知物体,还需要能处理训练过程中被忽视的未知物体。针对已知和未知物体的检测混淆、密集未知目标和小目标遗漏等问题,提出了一种新的基于偏移过滤和未知特征强化的开放世界目标检测器(offset filter and unknown-feature reinforcement for open world object detection,OFUR-OWOD)。首先设计一个未知类特征强化(unknown class feature reinforcement,UCFR)模块,通过自适应未知对象得分的方法来强化未知类目标特征,进而提高模型对未知类对象的训练准确度。然后,将重叠框偏移过滤器(overlapping box offset filter,OBOF)应用于目标预测框,根据目标位置和大小,获得不同偏移得分,以过滤冗余未知框。通过丰富实验证明,该方法在COCO-OOD和COCO-Mix上优于现有一些最先进的方法。OWOD is a computer vision challenge focusing on real-world environments,requiring the detection of both labeled known objects and previously unseen unknown objects during training.This paper proposed a novel open world object detector based on OFUR-OWOD to address the problems of confusing the detection of known and unknown objects,dense unknown objects and small object omission.Firstly,it designed an unknown class feature reinforcement(UCFR)module to reinforce the unknown object features by means of adaptive unknown object scoring,which in turn improved the training accuracy of the model for unknown objects.Then,it applied overlapping box offset filter(OBOF)to the proposals to obtain different offset scores according to the object position and size in order to filter redundant unknown object boxes.Extensive experiments demonstrate that this method outperforms some of the state-of-the-art methods on COCO-OOD and COCO-Mix.

关 键 词：开放世界目标检测 自适应得分 未知目标特征强化 偏移过滤 

分 类 号：P315.69[天文地球—地震学]