[1]姜孟超,范灵毓,李硕豪*.基于注意力双线性池化的细粒度舰船识别[J].计算机技术与发展,2022,32(08):66-70.[doi:10. 3969 / j. issn. 1673-629X. 2022. 08. 011]
　JIANG Meng-chao,FAN Ling-yu,LI Shuo-hao*.Weakly Supervised Fine-grained Natural Scene Ship Recognition viaAttention Bilinear Pooling[J].,2022,32(08):66-70.[doi:10. 3969 / j. issn. 1673-629X. 2022. 08. 011]
点击复制

基于注意力双线性池化的细粒度舰船识别()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
32
期数:
2022年08期
页码:
66-70
栏目:
图形与图像
出版日期:
2022-08-10

文章信息/Info

Title:
Weakly Supervised Fine-grained Natural Scene Ship Recognition viaAttention Bilinear Pooling
文章编号:
1673-629X(2022)08-0066-05
作者:
姜孟超1 ; 范灵毓2 ; 李硕豪3*
1. 军事科学院 战略评估咨询中心,北京 100091;
2. 96962 部队,北京 102206;
3. 国防科技大学 信息系统工程重点实验室,湖南 长沙 410073
Author(s):
JIANG Meng-chao1 ; FAN Ling-yu2 ; LI Shuo-hao3*
1. Consulting Center for Strategic Assessment,Academy of Military Science,Beijing 100091,China;
2. Unit 96962,Beijing 102206,China;
3. Key Laboratory of Information Systems Engineering,National University of Defense Technology,Changsha 410073,China
关键词:
弱监督学习; 细粒度图像识别; 通道注意力机制; 空间注意力机制; 双线性池化
Keywords:
weakly supervised learning;  fine - grained image recognition;  channel attention mechanism;  spatial attention mechanism; bilinear pooling
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 08. 011
摘要:
针对当前细粒度图像识别的模型结构复杂且只能学习到单一判别性特征的问题,对一种弱监督学习下基于注意力双线性池化的细粒度舰船识别方法进行研究。 该方法首先将通道注意力机制和空间注意力机制与卷积神经网络相结合,在没有监督信息的条件下分别提取图像的深度通道特征和深度空间特征。 然后通过双线性池化操作对提取到的深度通道特征和深度空间特征进行特征融合,使得通道特征和空间特征形成关联和交互,从而使网络能够学习到更丰富的图像局部特征。 最后再将学习到的局部特征和深度神经网络提取到的全局特征进行拼接,利用全连接层得到最终的图像融合特征用于舰船图像的细粒度分类。 针对当前缺少自然场景下的舰船数据集问题,进行了相关舰船图像数据的收集工作,建立了针对自然场景舰船细粒度检测的数据集,并在该数据集上进行了训练和测试,该模型的识别准确率可以达到91. 3% 。
Abstract:
Aiming at the problem that the current model structure of fine-grained image recognition is complex and can only learn a singlediscriminant feature, a fine - grained ship recognition method based on attention bilinear pooling under weak supervised learning isstudied. Firstly,the channel attention mechanism and spatial attention mechanism are combined with convolution neural network toextract the depth channel features and depth space features of the image without supervision information. Then, the extracted depthchannel features and depth spatial features are fused through bilinear pooling operation,so that the channel features and spatial featuresform association and interaction,making the network can learn more abundant image local features. Finally,the learned local features andthe global features extracted by depth neural network are spliced, and the final image fusion features are obtained by using the fullconnection layer for fine-grained classification of ship images. In view of the current lack of ship data sets in open natural scenes,wecollect relevant ship image data and establish a data set for fine-grained ship detection in natural scenes. Through training and testing onthe data set,the recognition accuracy of such model reaches 91. 3% .

相似文献/References:

[1]范子祎,杨 欢.LLP-AAE 算法在金融风险识别领域的应用[J].计算机技术与发展,2021,31(增刊):151.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 030]
　FAN Zi-yi,YANG Huan.Adversarial Auto-encoding for Learning from Label Proportions on Financial Risk Identification[J].,2021,31(08):151.[doi:10. 3969 / j. issn. 1673-629X. 2021. S. 030]
[2]杨 虹,范 勇.一种基于区分区域定位的细粒度图像识别方法[J].计算机技术与发展,2023,33(11):169.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 025]
　YANG Hong,FAN Yong.A Fine Grain Image Recognition Method Based on Distinguishable Region Location[J].,2023,33(08):169.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 025]

更新日期/Last Update: 2022-08-10