基于动态自适应通道注意力特征融合的小目标检测  

Small object detection based on dynamic adaptive channel attention feature fusion

作　　者：吴迪[1,2] 赵品懿[1] 甘升隆 沈学军 万琴[1,2] WU Di;ZHAO Pinyi;GAN Shengong;SHEN Xuejun;WAN Qin(Institute of Electrical and Information Engineering,Hunan Institute of Engineering,Xiangtan 411100,China;National Engineering Research Center of RVC,Hunan University,Changsha 410082,China)

机构地区：[1]湖南工程学院电气与信息工程学院,湘潭411100 [2]湖南大学机器人视觉感知与控制技术国家工程研究中心,长沙410082

出　　处：《电子科技大学学报》2025年第2期221-232,共12页Journal of University of Electronic Science and Technology of China

基　　金：国家重点研发计划(2020YFB1713600);国家自然科学基金(62476084);湖南省教育厅重点项目(24A0528);湖南省自然科学基金(2022JJ30198);湖南省研究生科研创新项目(YC202213)。

摘　　要：针对小目标检测中卷积操作导致检测特征缺失和不同尺度语义隔阂的问题,提出一种基于动态自适应通道注意力特征融合的小目标检测方法。1)提出一种多尺度三角动态颈(Tri-Neck)网络结构,用于融合多尺度特征语义隔阂及弥补小目标特征缺失的问题。2)提出一种分组批量动态自适应通道注意力模块,增强弱语义小目标特征同时抑制无用信息,且在动态自适应通道注意力模块中设计新的激活函数和交并比损失函数,提升通道注意力表征能力。3)采用ResNet50作为骨干网络依次连接特征金字塔网络和Tri-Neck网络。实验结果表明,该方法在Pascal Voc 2007、Pascal Voc 2012上比YOLOv8算法mAP分别提升5.3%和6.2%,在MS COCO 2017数据集上AP和AP_S分别提升1.6%和2%,在SODA-D数据集上比YOLOv8算法AP提升0.9%。In order to solve the feature missing and different scales features semantics gaps problem causedby convolution operations in small object detection,a small object detection method based on dynamic adaptivechannel attention feature fusion is proposed in this paper.Firstly,a Tri-Neck network structure is introduced toaddress the semantic gaps and feature deficiency in small object detection across multiple scales.Secondly,adynamic adaptive channel attention module is proposed to enhance weak semantic features of small objects whilesuppressing irrelevant information.Additionally,new activation functions and intersection-over-union lossfunctions are designed within the dynamic adaptive channel attention module to improve channel attentionrepresentation capability.Finally,the ResNet50 backbone network is utilized,connecting the feature pyramidnetwork and the Tri-Neck network sequentially.Experimental results on the Pascal VOC 2007 and Pascal VOC2012 datasets demonstrate performance improvements of 5.3%and 6.2%respectively,while on the MS COCO2017 dataset,the proposed algorithm shows enhancements in overall performance and small object detectionperformance by 1.6%and 2%respectively,and on the SODA-D dataset,our proposed algorithm demonstratessuperior performance compared to the suboptimal algorithm AP,resulting in a 0.9%improvement in overallaccuracy.

关 键 词：小目标检测 多尺度融合特征 特征金字塔 动态通道注意力 交并比损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TH701[自动化与计算机技术—计算机科学与技术]