第16卷第6期 智能系统学报 Vol.16 No.6 2021年11月 CAAI Transactions on Intelligent Systems Nov.2021 D0:10.11992/tis.202012029 网络出版地址：https:/ns.cnki.net/kcms/detail/23.1538.TP.20210903.1028.002.html 双向特征融合与注意力机制结合的目标检测 赵文清2，杨盼盼 (1.华北电力大学控制与计算机工程学院，河北保定071003；2.复杂能源系统智能计算教育部工程研究中心， 河北保定071003) 摘要：目标检测使用特征金字塔检测不同尺度的物体时，忽略了高层信息和低层信息之间的关系，导致检测 效果差：此外，针对某些尺度的目标，检测中容易出现漏检。本文提出双向特征融合与注意力机制结合的方法 进行目标检测。首先，对SSD(single shot multibox detector)模型深层特征层与浅层特征层进行特征融合，然后将 得到的特征与深层特征层进行融合。其次，在双向融合中加入了通道注意力机制，增强了语义信息。最后，提 出了一种改进的正负样本判定策略，降低目标的漏检率。将本文提出的算法与当前主流算法在VOC数据集上 进行了比较，结果表明，本文提出的算法在对目标进行检测时，目标平均准确率有较大提高。 关键词：特征金字塔；双向融合；特征提取；SNet注意力机制：样本；语义信息；目标检测；深度学习 中图分类号：TP391 文献标志码：A文章编号：1673-4785(2021)06-1098-08 中文引用格式：赵文清，杨盼盼.双向特征融合与注意力机制结合的目标检测.智能系统学报，2021,16(6)：1098-1105. 英文引用格式：ZHAO Wenqing,YANG Panpan.Target detection based on bidirectional feature fusion and an attention mechan- ismJI.CAAI transactions on intelligent systems,2021,16(6):1098-1105. Target detection based on bidirectional feature fusion and an attention mechanism ZHAO Wenqing,YANG Panpan' (1.School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China;2.Engineering Re- search Center of the Ministry of education for Intelligent Computing of Complex Energy System,Baoding 071003,China) Abstract:When using a feature pyramid to detect objects of different dimensions,the relationship between high-and low-level information is ignored,resulting in a poor detection effect;in addition,for targets of a certain scale,detection is easily missed.In this paper,a method combining bidirectional feature fusion and an attention mechanism is proposed for target detection.First,the deep and shallow feature layers of the single-shot multibox detector(SSD)model are fused,then the obtained features are fused with the deep feature layer.Second,the channel attention mechanism is ad- ded to the two-way fusion to enhance semantic information.Finally,an improved positive and negative sample decision strategy is proposed to reduce the target misdetection rate.The algorithm proposed in this paper is compared with the current mainstream algorithms in the VOC dataset.The results show that the average accuracy of the proposed al- gorithm is greatly improved when detecting targets. Keywords:feature pyramid;bidirectional fusion;feature extraction;SeNet attention mechanism;sample;semantic in- formation;target detection;deep learning 目标检测是计算机视觉领域的重要研究方向。现阶段的目标检测方法主要有2种：一种是 基于分类的两阶段法，另一种是基于回归的单阶 收稿日期：2020-12-17.网络出版日期：2021-09-03. 段法。20l4年Girshick等四首次提出R-CNN(re- 基金项目：河北省自然科学基金项目(F2021502013):中央高校 基本科研业务费面上项目(2020MS153,2021PT018). gion convolutional neural networ,R-CNN)使用选择 通信作者：赵文清.E-mail:zhaowenqing(@ncepu.edu..cn 性搜索生成的区域建议。该算法相比传统算法DOI: 10.11992/tis.202012029 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20210903.1028.002.html 双向特征融合与注意力机制结合的目标检测 赵文清1,2，杨盼盼1 （1. 华北电力大学 控制与计算机工程学院，河北 保定 071003; 2. 复杂能源系统智能计算教育部工程研究中心， 河北 保定 071003） 摘 要：目标检测使用特征金字塔检测不同尺度的物体时，忽略了高层信息和低层信息之间的关系，导致检测 效果差；此外，针对某些尺度的目标，检测中容易出现漏检。本文提出双向特征融合与注意力机制结合的方法 进行目标检测。首先，对 SSD(single shot multibox detector) 模型深层特征层与浅层特征层进行特征融合，然后将 得到的特征与深层特征层进行融合。其次，在双向融合中加入了通道注意力机制，增强了语义信息。最后，提 出了一种改进的正负样本判定策略，降低目标的漏检率。将本文提出的算法与当前主流算法在 VOC 数据集上 进行了比较，结果表明，本文提出的算法在对目标进行检测时，目标平均准确率有较大提高。 关键词：特征金字塔；双向融合；特征提取；SeNet 注意力机制；样本；语义信息；目标检测；深度学习 中图分类号：TP391 文献标志码：A 文章编号：1673−4785(2021)06−1098−08 中文引用格式：赵文清, 杨盼盼. 双向特征融合与注意力机制结合的目标检测 [J]. 智能系统学报, 2021, 16(6): 1098–1105. 英文引用格式：ZHAO Wenqing, YANG Panpan. Target detection based on bidirectional feature fusion and an attention mechan￾ism[J]. CAAI transactions on intelligent systems, 2021, 16(6): 1098–1105. Target detection based on bidirectional feature fusion and an attention mechanism ZHAO Wenqing1,2 ，YANG Panpan1 (1. School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China; 2. Engineering Re￾search Center of the Ministry of education for Intelligent Computing of Complex Energy System, Baoding 071003, China) Abstract: When using a feature pyramid to detect objects of different dimensions, the relationship between high- and low-level information is ignored, resulting in a poor detection effect; in addition, for targets of a certain scale, detection is easily missed. In this paper, a method combining bidirectional feature fusion and an attention mechanism is proposed for target detection. First, the deep and shallow feature layers of the single-shot multibox detector (SSD) model are fused, then the obtained features are fused with the deep feature layer. Second, the channel attention mechanism is ad￾ded to the two-way fusion to enhance semantic information. Finally, an improved positive and negative sample decision strategy is proposed to reduce the target misdetection rate. The algorithm proposed in this paper is compared with the current mainstream algorithms in the VOC dataset. The results show that the average accuracy of the proposed al￾gorithm is greatly improved when detecting targets. Keywords: feature pyramid; bidirectional fusion; feature extraction; SeNet attention mechanism; sample; semantic in￾formation; target detection; deep learning 目标检测是计算机视觉领域的重要研究方 向。现阶段的目标检测方法主要有 2 种：一种是 基于分类的两阶段法，另一种是基于回归的单阶 段法。2014 年 Girshick 等 [1] 首次提出 R-CNN(re￾gion convolutional neural networ，R-CNN) 使用选择 性搜索[2] 生成的区域建议。该算法相比传统算法 收稿日期：2020−12−17. 网络出版日期：2021−09−03. 基金项目：河北省自然科学基金项目 (F2021502013)；中央高校 基本科研业务费面上项目 (2020MS153，2021PT018). 通信作者：赵文清. E-mail：zhaowenqing@ncepu.edu.cn. 第 16 卷第 6 期 智 能 系 统 学 报 Vol.16 No.6 2021 年 11 月 CAAI Transactions on Intelligent Systems Nov. 2021