检索规则说明:AND代表“并且”;OR代表“或者”;NOT代表“不包含”;(注意必须大写,运算符两边需空一格)
检 索 范 例 :范例一: (K=图书馆学 OR K=情报学) AND A=范并思 范例二:J=计算机应用与软件 AND (U=C++ OR U=Basic) NOT M=Visual
作 者:尹芹[1] 方晖[1] 王金东[1] 王侃[1] 晏天文 霍智勇[2] YIN Qin;FANG Hui;WANG Jindong;WANG Kan;YAN Tianwen;HUO Zhiyong(Multimedia Video Products Department,ZTE Corporation,Nanjing 210023,China;School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
机构地区:[1]中兴通讯股份有限公司多媒体视讯产品部,江苏南京210000 [2]南京邮电大学通信与信息工程学院,江苏南京210003
出 处:《南京邮电大学学报(自然科学版)》2022年第4期69-74,共6页Journal of Nanjing University of Posts and Telecommunications:Natural Science Edition
基 金:中兴通讯研究基金资助项目。
摘 要:小目标检测是计算机视觉领域具有挑战性的问题。空间注意力和通道注意力机制的使用提高了目标检测网络的均值平均精度,但捕获小物体上下文信息的能力仍然有限,并且在小目标和大中型目标的检测精度上存在差距,难以感知小物体的位置。算法构建了一种基于通道自注意力机制(Channel Self-Attention, CSA)的算法模块,将输入特征映射压缩后,运用自注意力机制建立特征通道间相关性,自适应地重新优化特征通道的响应,提升了捕获小物体远距离上下文信息的能力,从而提高了对小目标的检测精度。实验结果表明,在几乎不增加计算成本的情况下,CSA块能够为现有目标检测网络带来性能改进。在PASCAL VOC2007数据集上,采用通道自注意力机制的RetinaNet的mAP值分别比原始RetinaNet的mAP值高3.11个百分点。使用通道自注意力机制的MobileNetv2比原始的MobileNetv2 mAP值提高3.05个百分点。Small object detection is a challenging problem in the field of computer vision. Some networks adopt the spatial attention and the channel attention mechanisms to improve the mean average accuracy of object detection, but their ability of capturing contextual information of small objects is still limited, and they cannot present the same detection accuracy on small objects as they do on large and medium-sized objects. The paper constructs an algorithm based on the channel self-attention mechanism(CSA). This algorithm deploys the self-attention mechanism after compressing the input feature map to establish the correlation between feature channels, and thus adaptively re-optimizes the feature channels. In this way, the ability of capturing long-distance context information of small objects is improved, and the detection accuracy for small objects is therefore increased. Experimental results show that the CSA block can help to improve current networks’ performance with little increased computational cost. On the PASCAL VOC2007 dataset, the mAP values of the RetinaNet with the channel self-attention mechanism are 3.11 percent higher than those of the original RetinaNet;and the mAP value of the MobileNetv2 with the channel self-attention is increased by 3.05 percent from that of the original MobileNetv2.
关 键 词:注意力机制 小目标检测 自注意力 通道注意力 空间注意力
分 类 号:TP391.41[自动化与计算机技术—计算机应用技术]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在链接到云南高校图书馆文献保障联盟下载...
云南高校图书馆联盟文献共享服务平台 版权所有©
您的IP:216.73.216.38