[1]林紫心,陈东方,王晓峰.基于改进Dense Teacher的半监督目标检测算法[J].计算机技术与发展,2024,34(12):148-156.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0252]
　LIN Zi-xin,CHEN Dong-fang,WANG Xiao-feng.Semi-supervised Object Detection Algorithm Based on Improved Dense Teacher[J].,2024,34(12):148-156.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0252]
点击复制

基于改进Dense Teacher的半监督目标检测算法()

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

卷:
34
期数:
2024年12期
页码:
148-156
栏目:
人工智能
出版日期:
2024-12-10

文章信息/Info

Title:
Semi-supervised Object Detection Algorithm Based on Improved Dense Teacher
文章编号:
1673-629X(2024)12-0148-09
作者:
林紫心1; 2; 陈东方1; 2; 王晓峰1; 2
1. 武汉科技大学 计算机科学与技术学院,湖北 武汉 430065;2. 智能信息处理与实时工业系统湖北省重点实验室,湖北 武汉 430065
Author(s):
LIN Zi-xin1; 2; CHEN Dong-fang1; 2; WANG Xiao-feng1; 2
1. School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;2. Hubei Provincial Key Laboratory of Intelligent Information Processing and Real-time Industrial Systems,Wuhan 430065,China
关键词:
无锚框目标检测器; 特征融合; 伪标签; 双分支处理策略; 半监督学习; 半监督目标检测
Keywords:
anchor-free object detector; feature fusion; pseudo-labels; double branch processing strategy; semi-supervised learning; semi-supervised object detection
分类号:
TP391.41
DOI:
10.20165/j.cnki.ISSN1673-629X.2024.0252
摘要:
在深度学习目标检测领域中,标注数据的复杂性对半监督目标检测技术的发展起到了推动作用。 近期,无锚框目标检测器在半监督目标检测中的应用逐渐增多。 然而,现有方法仍面临特征表达不足和伪标签质量参差不齐的挑战。 为解决这一问题,该文提出了一种基于优化 Dense Teacher 的半监督目标检测算法。 首先,通过设计通道信息增强的特征金字塔(CIE-FPN),优化了特征图的感受野和特征融合,旨在捕获更丰富的全局特征信息。 其次,采用伪标签双分支处理策略,以解决分类与检测任务在伪标签需求上的不一致性,从而提高伪标签的可靠性。 实验结果表明,在 COCO 数据集仅使用 10% 有标注数据的条件下,与基线网络 Dense Teacher 相比,提出的改进算法在提升无锚框半监督目标检测性能方面取得了明显效果。
Abstract:
The complexity of annotation in deep learning - based object detection has spurred advancements in semi - supervised target detection techniques. Recently,anchor-free object detectors have been applied in semi-supervised object detection,however existing methods face problems of inadequate feature representation and low quality of pseudo labels. For this purpose,we propose an semi-supervised object detection algorithm based on improved Dense Teacher. Firstly, a feature pyramid ( CIE - FPN) based on channel information enhancement is designed to optimize the receptive field and feature fusion of feature maps,so as to capture more global feature information. Secondly, a pseudo - label double - branch processing strategy is adopted to solve the inconsistency between classification and detection tasks,and then improve the robustness of pseudo-labels. Experimental results on the COCO dataset,with only 10% of data annotated,demonstrate significant improvements in anchor-free semi-supervised object detection performance compared to the baseline Dense Teacher network.

相似文献/References:

[1]周伟 武港山.基于显著图的花卉图像分类算法研究[J].计算机技术与发展,2011,(11):15.
　ZHOU Wei,WU Gang-shan.Research on Saliency Map Based Flower Image Classification Algorithm[J].,2011,(12):15.
[2]黎粤华,单磊,田仲富,等. 基于多特征融合的视频烟雾检测[J].计算机技术与发展,2016,26(01):129.
　LI Yue-hua,SHAN Lei,TIAN Zhong-fu,et al. Video Smoke Detection Based on Multi Feature Fusion Technology[J].,2016,26(12):129.
[3]刘加运,李玉惠,李勃,等. 一种多维特征融合的车辆对象同一性匹配方法[J].计算机技术与发展,2016,26(04):167.
　LIU Jia-yun,LI Yu-hui,LI Bo,et al. A Vehicle Object Identity Matching Method of Multidimensional Feature Combination[J].,2016,26(12):167.
[4]陈浩翔,蔡建明,刘铿然,等. 手写数字深度特征学习与识别[J].计算机技术与发展,2016,26(07):19.
　CHEN Hao-xiang,CAI Jian-ming,LIU Keng-ran,et al. Deep Learning and Recognition of Handwritten Numeral Features[J].,2016,26(12):19.
[5]张雅倩,曾卫明,石玉虎.基于特征融合与稀疏表示的人耳识别[J].计算机技术与发展,2017,27(12):7.
　ZHANG Ya-qian,ZENG Wei-min,SHI Yu-hu.Ear Recognition Based on Feature Fusion and Sparse Representation[J].,2017,27(12):7.
[6]谭程午,夏利民,王 嘉.基于融合特征的群体行为识别[J].计算机技术与发展,2018,28(01):17.[doi:10.3969/ j. issn.1673-629X.2018.01.004]
　TAN Cheng-wu,XIA Li-min,WANG Jia.Recognition of Human Group Action Based on Fusion Features[J].,2018,28(12):17.[doi:10.3969/ j. issn.1673-629X.2018.01.004]
[7]王 敏,陈立潮,曹建芳,等.Hadoop 下自适应随机权值多特征融合图像分类[J].计算机技术与发展,2018,28(11):30.[doi:10.3969/ j. issn.1673-629X.2018.11.007]
　WANG Min,CHEN Li-chao,CAO Jian-fang,et al.Multi-feature Fusion Image Classification of Adaptive Random Weight Based on Hadoop[J].,2018,28(12):30.[doi:10.3969/ j. issn.1673-629X.2018.11.007]
[8]韩欣欣,叶奇玲.基于 SIFT 和 HOG 特征融合的人体行为识别方法[J].计算机技术与发展,2019,29(06):71.[doi:10. 3969 / j. issn. 1673-629X. 2019. 06. 015]
　HAN Xin-xin,YE Qi-ling.Human Action Recognition Based on Feature Fusion of SIFT and HOG[J].,2019,29(12):71.[doi:10. 3969 / j. issn. 1673-629X. 2019. 06. 015]
[9]宋相法,吕 明.融合三维骨架和深度图像特征的人体行为识别[J].计算机技术与发展,2019,29(07):55.[doi:10. 3969 / j. issn. 1673-629X. 2019. 07. 011]
　SONG Xiang-fa,LYU Ming.Human Activity Recognition Based on Fusing 3D Skeleton and Depth Image Feature[J].,2019,29(12):55.[doi:10. 3969 / j. issn. 1673-629X. 2019. 07. 011]
[10]王泽泓,刘厚泉.基于迁移学习与自适应特征融合的建筑物识别[J].计算机技术与发展,2019,29(12):40.[doi:10. 3969 / j. issn. 1673-629X. 2019. 12. 007]
　WANG Ze-hong,LIU Hou-quan.Building Recognition Based on Transfer Learning and Adaptive Feature Fusion[J].,2019,29(12):40.[doi:10. 3969 / j. issn. 1673-629X. 2019. 12. 007]

更新日期/Last Update: 2024-12-10