计算机科学

计算机科学 ›› 2026, Vol. 53 ›› Issue (6A): 250700088-7.doi: 10.11896/jsjkx.250700088

• 图像处理&多媒体技术 • 上一篇    下一篇

基于分阶段训练策略与多尺度特征融合的目标检测方法

瞿洁武1, 路新喜2, 孙健1, 柳燕1, 高玲1, 徐彬彬1   

  1. 1 北京智网数科技术有限公司 北京 102200
    2 北京航空航天大学软件学院 北京 100191
  • 出版日期:2026-06-16 发布日期:2026-06-12
  • 通讯作者: 路新喜(lxx@buaa.edu.cn)
  • 作者简介:(ford.qu@163.com)

Object Detection Method Based on Phased Training Strategy and Multi-scale Feature Fusion

QU Jiewu1, LU Xinxi2, SUN Jian1, LIU Yan1, GAO Ling1, XU Binbin1   

  1. 1 Pipechina Digital Co.,Ltd.,Beijing 102200,China
    2 School of Software,Beihang University,Beijing 100191,China
  • Published:2026-06-16 Online:2026-06-12
  • About author:QU Jiewu,born in 1985.His main research interests include network communication and artificial intelligence.
    LU Xinxi,born in 1978,Ph.D,associate professor,master's supervisor.His main research interests include intelligent software and artificial intelligence.

PDF (PC)

35

摘要： 针对DETR(Detection Transformer)系列目标检测方法在推理阶段存在计算瓶颈、难以兼顾实时性与精度的问题,提出了一种基于分阶段训练策略与多尺度特征融合的改进方法。具体而言,通过简化DETR的多层编码器结构降低计算复杂度,并采用分阶段训练策略提升特征表达能力和模型收敛速度。第一阶段采用一对多标签匹配获取高质量二维多尺度特征,第二阶段冻结第一阶段的网络权重,并引入并行注意力卷积融合模块进一步细化特征。实验结果表明,所提方法在COCO数据集上较基线模型实现了5倍的推理速度提升,并带来了1.5个百分点的AP增益,有效缓解了DETR在推理阶段效率低下的问题;在BitVehicle数据集上也取得了1.4个百分点的AP提升。

关键词: 目标检测, 并行注意力卷积, 分阶段训练策略, 多尺度特征融合

Abstract: To overcome these limitations,such as the computational bottlenecks and the challenge of balancing real-time perfor-mance with accuracy in the DETR(Detection Transformer) family of object detection methods during inference,this paper proposes an enhanced approach that combines a phased training strategy with multi-scale feature fusion.Specifically,the multi-layer encoder structure of DETR is simplified to reduce computational complexity,while the phased training strategy improves feature representation and accelerates model convergence.In the first phase,one-to-many label matching is adopted to obtain high-quality two-dimensional multi-scale features.In the second phase,the weights from the first phase are frozen,and a parallel attention-convolutional fusion module is introduced to further refine the features.Experimental results demonstrate that the proposed method achieves a 5× increase in inference speed and a 1.5-point AP gain over the baseline model on the COCO dataset,effectively alleviating DETR's inference inefficiency.In addition,it yields a 1.4-point AP improvement on the BitVehicle dataset.

Key words: Object detection, Parallel attention convolution, Phased training strategy, Multi-scale feature fusion

中图分类号: 

  • TP311
[1] CARION N,MASSA F,SYNNAEVE G,et al.End-to-end object detection with transformers[C]//European Conference on Computer Vision.Cham:Springer International Publishing,2020:213-229.
[2] ZHU X,SU W,LU L,et al.Deformable detr:Deformable transformers for end-to-end object detection[J].arXiv:2010.04159,2020.
[3] LIU S,LI F,ZHANG H,et al.Dab-detr:Dynamic anchor boxes are better queries for detr[J].arXiv:2201.12329,2022.
[4] ZHANG H,LI F,LIU S,et al.Dino:Detr with improved denoi-sing anchor boxes for end-to-end object detection[J].arXiv:2203.03605,2022.
[5] LI F,ZENG A,LIU S,et al.Lite detr:An interleaved multi-scale encoder for efficient detr[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2023:18558-18567.
[6] REN S,HE K,GIRSHICK R,et al.Faster R-CNN:Towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2016,39(6):1137-1149.
[7] HE K,GKIOXARI G,DOLLÁR P,et al.Mask r-cnn[C]//Proceedings of the IEEE International Conference on Computer Vision.2017:2961-2969.
[8] CAI Z,VASCONCELOS N.Cascade r-cnn:Delving into highquality object detection[C]//Proceedings of the IEEE Confe-rence on Computer Vision and Pattern Recognition.2018:6154-6162.
[9] WANG C Y,BOCHKOVSKIY A,LIAO H Y M.YOLOv7:Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]//Proceedings of the IEEE/CVF Confe-rence on Computer Vision and Pattern Recognition.2023:7464-7475.
[10] LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for dense object detection[C]//Proceedings of the IEEE International Conference on Computer Vision.2017:2980-2988.
[11] WU J,ZHAO C.Small Object Detection Method Based on Improved DETR Algorithm[J/OL].Computer Applications,2025.
[12] MENG D,CHEN X,FAN Z,et al.Conditional detr for fasttraining convergence[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.2021:3651-3660.
[13] YAO Z,AI J,LI B,et al.Efficient detr:improving end-to-endobject detector with dense prior[J].arXiv:2104.01318,2021.
[14] WANG Y,ZHANG X,YANG T,et al.Anchor detr:Query design for transformer-based detector[C]//Proceedings of the AAAI Conference on Artificial Intelligence.2022:2567-2575.
[15] ZHANG D P,WEI Y Y,HE S J,et al.Feature Fusion and Inter-layer Transmission:An Improved Object Detection Method Based on Anchor DETR[J].Journal of Graphics,2024,45(5):968-978.
[16] LI F,ZHANG H,LIU S,et al.Dn-detr:Accelerate detr training by introducing query denoising[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2022:13619-13627.
[17] ZHENG D,DONG W,HU H,et al.Less is more:Focus attention for efficient detr[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.2023:6674-6683.
[18] ZHAO Y,LV W,XU S,et al.Detrs beat yolos on real-time object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2024:16965-16974.
[19] GUAN Y,LIAO S,YANG W.AParC-DETR:accelerate DETR training by introducing adaptive position-aware circular convolution[J].The Visual Computer,2025,41(2):1319-1333.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市两江新区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发