[1]张永福,宋海林.基于跳跃特征金字塔的域适应目标检测模型[J].计算机技术与发展,2022,32(09):28-35.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 005]
　ZHANG Yong-fu,SONG Hai-lin.Skip Feature Pyramid Based Domain Adapted Model for Object Detection[J].,2022,32(09):28-35.[doi:10. 3969 / j. issn. 1673-629X. 2022. 09. 005]
点击复制

基于跳跃特征金字塔的域适应目标检测模型()
分享到：

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

卷:
32
期数:
2022年09期
页码:
28-35
栏目:
媒体计算
出版日期:
2022-09-10

文章信息/Info

Title:
Skip Feature Pyramid Based Domain Adapted Model for Object Detection
文章编号:
1673-629X(2022)09-0028-08
作者:
张永福; 宋海林
陕西师范大学 计算机科学学院,陕西 西安 710119
Author(s):
ZHANG Yong-fu; SONG Hai-lin
School of Computer Science,Shaanxi Normal University,Xi’an 710119,China
关键词:
遥感图像; 目标检测; 特征金字塔; 跳跃连接; 域适应
Keywords:
remote sensing image; object detection; feature pyramid; skip connection; domain adaptation
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 09. 005
摘要:
针对训练数据和测试数据来源不同,特征分布差异较大,以及目标大小不一时,深度目标检测模型性能下降明显的问题,提出基于跳跃特征金字塔的域适应遥感图像目标检测模型。 首先,在目标检测网络中增加域适应部分,将训练数据作为源域,测试数据作为目标域,通过对抗训练的方式,对具有不同特征分布的两个域,关注其差异,同时提取源域和目标域数据的特征,减小遥感图像由光照、角度等不同造成的源域和目标域图像在图像级的域偏移对目标检测的影响,提升检测模型的推广性能。 其次,提出跳跃特征金字塔结构,通过特征上采样,以及同层连接、相隔层间的跳跃连接与特征融合,来增强特征图的细节信息和语义信息,以提高模型对不同尺度目标的检测精度。 最后,使用区域推荐网络在多个不同分辨率的特征图上提取候选区域,检测不同尺度的目标。 所提模型在 NWPUVHR-10 数据集上检测精度达到了 98. 2% 、误检率为 5. 4% 、漏检率为 8. 3% ;在 RSOD - DATA 的低亮度数据集上检测精度达到了 62% ,误检率、漏检率为 18. 2% 和18郾 5% 。 与其他模型相比,所提模型的性能有明显提升,具有更好的推广性。
Abstract:
In order to solve the problem that the detection performance of the deep object detection model drops significantly when the feature distribution differ greatly between training data and testing data, and when objects vary in size, a skip feature pyramid based domain adapted model for remote sensing image object detection is proposed. Firstly,the domain adaptation part is added into the objectdetection network. Regarding training data as the source domain and testing data as the target domain and focusing on their differences,the network can extract robust deep convolution features for both source and target domains with different feature distributions,through the adversarial training. This reduces the influence of the deviation between source and target domain image due to different illumination and angle in image acquisition, and improves the generalization of the model. Secondly,a skip feature pyramid structure is proposed to enhance the detail information and semantic information of the feature map by feature upsample,connection and jump connection between the same layer and the nonadjacent layer,and feature fusion,so as to improve the detection accuracy for different scale objects. Finally,the region proposal network is used to extract candidate regions on multiple feature maps with different resolutions to detect targets of different scales. The experimental results show that the detection accuracy on the NWPUVHR-10 dataset is 98. 2% ,the false detection rate is 5. 4% ,and the missed detection rate is? 8. 3% . The accuracy on the RSOD-DATA low-brightness dataset reaches 62% ,and the false detection rate and missed detection rate are? ?18. 2% and 18. 5% respectively. Compared with the other models,the performance of the proposed model has improved significantly with better generalization.

相似文献/References:

[1]蒋璐璐 王适 王宝成 李慧敏 李鑫慧.一种改进的标记分水岭遥感图像分割方法[J].计算机技术与发展,2010,(01):36.
　JIANG Lu-lu,WANG Shi,WANG Bao-cheng,et al.Segmentation of Remote Sensing Image Based on an Improved Labeling Watershed Algorithm[J].,2010,(09):36.
[2]刘晓明 李毓蕙 高燕 郑华强.基于目标区域清晰显示的H.264编码策略[J].计算机技术与发展,2010,(06):29.
　LIU Xiao-ming,LI Yu-hui,GAO Yan,et al.A Coding Strategy of H.264 Based on High-definition Display of Target Region[J].,2010,(09):29.
[3]刘翔 吴谨 祝愿博 康晓晶.基于视频序列的目标检测与跟踪技术研究[J].计算机技术与发展,2009,(11):179.
　LIU Xiang,WU Jin,ZHU Yuan-bo,et al.A Study of Object Detecting and Tracking Based on Video Sequences[J].,2009,(09):179.
[4]赵艳飞 高清维 卢一相.基于多尺度Retinex算法的遥感图像增强[J].计算机技术与发展,2008,(02):70.
　ZHAO Yan-fei,GAO Qing-wei,LU Yi-xiang.Remote Sensing Image Enhancement Based on Multi- Scale Retinex Algorithm[J].,2008,(09):70.
[5]陈昌生 许文波 赵健.基于ArcSDE的卫星遥感影像数据库研究[J].计算机技术与发展,2010,(07):80.
　CHEN Chang-sheng,XU Wen-bo,ZHAO Jian.Research on Remote Sensing Image Database Based on ArcSDE[J].,2010,(09):80.
[6]李晓飞 马大玮 范小麟 胡焰智.基于数学形态学的遥感图像多感兴趣区域提取[J].计算机技术与发展,2007,(12):56.
　LI Xiao-fei,MA Da-wei,FAN Xiao-lin,et al.Extraction of ROIs Based on Mathematical Morphology toward Applications on Remote Sense Image Processing[J].,2007,(09):56.
[7]曙光 张超 蔡则苏.基于改进的混合高斯模型的目标检测方法[J].计算机技术与发展,2012,(07):60.
　SHU Guang,ZHANG Chao,CAI Ze-su.Target Detection Method Based on Improved Gaussian Mixture Model[J].,2012,(09):60.
[8]刘洁,李目,周少武.一种混沌混合粒子群优化RBF神经网络算法[J].计算机技术与发展,2013,(08):181.
　LIU Jie[],LI Mu[],ZHOU Shao-wu[].An Algorithm of Chaotic Hybrid Particle Swarm Optimization Based on RBF Neural Network[J].,2013,(09):181.
[9]董敏[],阎镇.基于CUDA的航天遥感图像实时快视系统的研究[J].计算机技术与发展,2014,24(06):32.
　DONG Min[][][],YAN Zhen[].Research on Aerospace Remote Sensing Images Real-time Quick-view System Based on CUDA[J].,2014,24(09):32.
[10]蒋翠清,孙富亮,吴艿芯. 基于相对欧氏距离的背景差值法视频目标检测[J].计算机技术与发展,2015,25(01):37.
　JIANG Cui-qing,SUN Fu-liang,WU Nai-xin. Video Object Detection of Background Subtraction Method Based on Relative Euclidean Distance[J].,2015,25(09):37.
[11]张永福,宋海林,班 越,等.融合特征的深度学习遥感图像目标检测模型[J].计算机技术与发展,2021,31(09):48.[doi:10. 3969 / j. issn. 1673-629X. 2021. 09. 009]
　ZHANG Yong-fu,SONG Hai-lin,BAN Yue,et al.Fusion Features Based Deep Learning Remote Sensing Image Target Detection Model[J].,2021,31(09):48.[doi:10. 3969 / j. issn. 1673-629X. 2021. 09. 009]
[12]薛继伟,孙宇锐.基于改进 YOLOv5 的光学遥感图像水坝检测研究[J].计算机技术与发展,2023,33(05):69.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 011]
　XUE Ji-wei,SUN Yu-rui.Research on Dam Detection of Optical Remote Sensing Image Based on Improved YOLOv5[J].,2023,33(09):69.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 011]

更新日期/Last Update: 2022-09-10