《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (8): 2652-2659.DOI: 10.11772/j.issn.1001-9081.2025070856
• 多媒体计算与计算机仿真 • 上一篇
收稿日期:2025-07-29
修回日期:2025-09-19
接受日期:2025-09-25
发布日期:2025-11-05
出版日期:2026-08-10
通讯作者:
李英顺
作者简介:黄麟(2001—),女(满族),吉林东丰人,硕士研究生,主要研究方向:计算机视觉基金资助:
Lin HUANG1, Yingshun LI2(
), Weiyan TONG1, Shuyuan ZHANG3, Zihan WANG4
Received:2025-07-29
Revised:2025-09-19
Accepted:2025-09-25
Online:2025-11-05
Published:2026-08-10
Contact:
Yingshun LI
About author:HUANG Lin, born in 2001, M. S. candidate. Her research interests include computer vision.Supported by:摘要:
针对遥感图像背景复杂、小目标众多且存在长尾型目标导致的检测精度低的问题,提出一种基于改进实时检测Transformer (RT-DETR)的遥感图像目标检测算法SLT-DETR (Small and Long-Tailed object-aware real-time DEtection TRansformer)。首先,采用多尺度特征增强金字塔(MFEP)优化原有的连续卷积特征融合,并引入OmniKernel进一步增强对小目标的检测能力;其次,引入内容引导的注意力对P3、P4和P5的输出特征进行融合,从而有效解决信息分布失衡问题;最后,使用FasterBlock模块实现模型的轻量化,并提升推理速度。实验结果表明,在自建遥感数据集上,SLT-DETR的平均精度均值(mAP)相较于RT-DETR提升了5.0%,模型参数量降低了1.9×106;相较于经典的YOLOv8m、YOLOv8l以及Swin Transformer模型,SLT-DETR在mAP上分别提高了5.7%、3.7%和10.9%。在RSOD和HIT-UAV公开数据集上,SLT-DETR在mAP上分别提高了1.1%和1.3%。可见,所提模型能够在保证轻量化的同时有效提升对小目标与长尾型目标的检测能力。
中图分类号:
黄麟, 李英顺, 佟维妍, 张树园, 王子涵. 基于改进实时检测Transformer的遥感图像目标检测算法[J]. 计算机应用, 2026, 46(8): 2652-2659.
Lin HUANG, Yingshun LI, Weiyan TONG, Shuyuan ZHANG, Zihan WANG. Remote sensing image object detection algorithm based on improved real-time detection Transformer[J]. Journal of Computer Applications, 2026, 46(8): 2652-2659.
| 类别 | 目标数 | 目标占比/% | 类别 | 目标数 | 目标占比/% |
|---|---|---|---|---|---|
| airport | 9 255 | 20.71 | oiltank | 1 230 | 2.75 |
| runway | 22 818 | 51.05 | playground | 171 | 0.38 |
| aircraft | 11 220 | 25.10 |
表1 各类型目标数及占比
Tab. 1 Number and proportion of each type of target
| 类别 | 目标数 | 目标占比/% | 类别 | 目标数 | 目标占比/% |
|---|---|---|---|---|---|
| airport | 9 255 | 20.71 | oiltank | 1 230 | 2.75 |
| runway | 22 818 | 51.05 | playground | 171 | 0.38 |
| aircraft | 11 220 | 25.10 |
| MFEP-OK | CGAF | FasterBlock | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | FLOPs/109 |
|---|---|---|---|---|---|---|---|---|
| × | × | × | 92.0 | 85.8 | 91.1 | 64.1 | 19.8 | 57.0 |
| √ | × | × | 93.6 | 88.4 | 93.8 | 71.0 | 20.5 | 65.2 |
| × | √ | × | 94.4 | 90.2 | 94.4 | 71.8 | 21.0 | 59.2 |
| × | × | √ | 89.8 | 85.4 | 90.8 | 66.7 | 16.8 | 49.5 |
| × | √ | √ | 96.5 | 88.2 | 93.6 | 69.4 | 17.3 | 51.8 |
| √ | × | √ | 94.6 | 86.3 | 92.8 | 69.2 | 17.4 | 57.7 |
| √ | √ | × | 92.7 | 92.6 | 96.6 | 73.7 | 21.0 | 67.5 |
| √ | √ | √ | 96.3 | 91.8 | 95.7 | 72.9 | 17.9 | 60.0 |
表2 自建数据集消融实验结果
Tab. 2 Results of ablation experiments on self-built dataset
| MFEP-OK | CGAF | FasterBlock | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | FLOPs/109 |
|---|---|---|---|---|---|---|---|---|
| × | × | × | 92.0 | 85.8 | 91.1 | 64.1 | 19.8 | 57.0 |
| √ | × | × | 93.6 | 88.4 | 93.8 | 71.0 | 20.5 | 65.2 |
| × | √ | × | 94.4 | 90.2 | 94.4 | 71.8 | 21.0 | 59.2 |
| × | × | √ | 89.8 | 85.4 | 90.8 | 66.7 | 16.8 | 49.5 |
| × | √ | √ | 96.5 | 88.2 | 93.6 | 69.4 | 17.3 | 51.8 |
| √ | × | √ | 94.6 | 86.3 | 92.8 | 69.2 | 17.4 | 57.7 |
| √ | √ | × | 92.7 | 92.6 | 96.6 | 73.7 | 21.0 | 67.5 |
| √ | √ | √ | 96.3 | 91.8 | 95.7 | 72.9 | 17.9 | 60.0 |
| 模型 | P | R | mAP@0.5 | mAP@0.5:0.95 |
|---|---|---|---|---|
| RT-DETR | 93.2 | 73.9 | 82.4 | 42.8 |
| RT-DETR+CGAF | 97.2 | 82.4 | 91.5 | 56.2 |
表3 引入CGAF模块前后的oiltank检测结果 (%)
Tab. 3 Detection results of oiltank before and after introduction of CGAF module
| 模型 | P | R | mAP@0.5 | mAP@0.5:0.95 |
|---|---|---|---|---|
| RT-DETR | 93.2 | 73.9 | 82.4 | 42.8 |
| RT-DETR+CGAF | 97.2 | 82.4 | 91.5 | 56.2 |
| 模型 | P | R | mAP@0.5 | mAP@0.5:0.95 |
|---|---|---|---|---|
| RT-DETR | 81.2 | 75.0 | 82.7 | 47.9 |
| RT-DETR+CGAF | 88.0 | 83.2 | 91.6 | 65.7 |
表4 引入CGAF模块前后的playground检测结果 (%)
Tab. 4 Detection results of playground before and after introduction of CGAF module
| 模型 | P | R | mAP@0.5 | mAP@0.5:0.95 |
|---|---|---|---|---|
| RT-DETR | 81.2 | 75.0 | 82.7 | 47.9 |
| RT-DETR+CGAF | 88.0 | 83.2 | 91.6 | 65.7 |
| 模型 | RSOD数据集 | HIT-UAV | FLOPs/109 | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | ||
| RT-DETR | 96.7 | 96.7 | 98.0 | 77.6 | 19.8 | 80.4 | 76.4 | 78.6 | 50.5 | 19.8 | 57.0 |
| SLT-DETR | 97.3 | 96.8 | 99.1 | 74.8 | 17.6 | 82.3 | 77.5 | 79.6 | 51.6 | 17.9 | 60.0 |
表5 RSOD和HIT-UAV数据集上的实验结果
Tab. 5 Experimental results on RSOD and HIT-UAV datasets
| 模型 | RSOD数据集 | HIT-UAV | FLOPs/109 | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | ||
| RT-DETR | 96.7 | 96.7 | 98.0 | 77.6 | 19.8 | 80.4 | 76.4 | 78.6 | 50.5 | 19.8 | 57.0 |
| SLT-DETR | 97.3 | 96.8 | 99.1 | 74.8 | 17.6 | 82.3 | 77.5 | 79.6 | 51.6 | 17.9 | 60.0 |
| 模型 | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | FLOPs/109 |
|---|---|---|---|---|---|---|
| YOLOv5m | 87.3 | 81.0 | 85.4 | 61.8 | 20.9 | 47.9 |
| YOLOv5l | 87.8 | 82.9 | 88.8 | 66.5 | 46.1 | 107.7 |
| YOLOv8m | 92.7 | 84.5 | 90.5 | 70.2 | 25.8 | 78.7 |
| YOLOv8l | 96.1 | 86.2 | 92.3 | 73.1 | 43.6 | 164.8 |
| RT-DETR+文献[ | 92.7 | 86.6 | 93.4 | 66.7 | 24.5 | 60.7 |
| DAttention | 94.1 | 83.9 | 90.2 | 66.3 | 20.5 | 57.0 |
| gConv+MFEP-OK+CGAF | 94.1 | 91.7 | 95.6 | 72.9 | 19.8 | 62.0 |
| PConv+MFEP-OK+CGAF | 92.0 | 91.3 | 94.4 | 71.4 | 15.3 | 54.3 |
| Swin Transformer | 87.3 | 78.6 | 86.3 | 61.2 | 36.3 | 97.0 |
| 文献[ | 94.5 | 88.1 | 94.3 | 70.7 | 21.2 | 75.1 |
| SLT-DETR | 96.3 | 91.8 | 95.7 | 72.9 | 17.9 | 60.0 |
表6 不同模型的性能对比
Tab. 6 Performance comparison of different models
| 模型 | P/% | R/% | mAP@0.5/% | mAP@0.5:0.95/% | 参数量/106 | FLOPs/109 |
|---|---|---|---|---|---|---|
| YOLOv5m | 87.3 | 81.0 | 85.4 | 61.8 | 20.9 | 47.9 |
| YOLOv5l | 87.8 | 82.9 | 88.8 | 66.5 | 46.1 | 107.7 |
| YOLOv8m | 92.7 | 84.5 | 90.5 | 70.2 | 25.8 | 78.7 |
| YOLOv8l | 96.1 | 86.2 | 92.3 | 73.1 | 43.6 | 164.8 |
| RT-DETR+文献[ | 92.7 | 86.6 | 93.4 | 66.7 | 24.5 | 60.7 |
| DAttention | 94.1 | 83.9 | 90.2 | 66.3 | 20.5 | 57.0 |
| gConv+MFEP-OK+CGAF | 94.1 | 91.7 | 95.6 | 72.9 | 19.8 | 62.0 |
| PConv+MFEP-OK+CGAF | 92.0 | 91.3 | 94.4 | 71.4 | 15.3 | 54.3 |
| Swin Transformer | 87.3 | 78.6 | 86.3 | 61.2 | 36.3 | 97.0 |
| 文献[ | 94.5 | 88.1 | 94.3 | 70.7 | 21.2 | 75.1 |
| SLT-DETR | 96.3 | 91.8 | 95.7 | 72.9 | 17.9 | 60.0 |
| [1] | 王长龙,冀鲸宇,赵月飞,等. 无人机遥感图像目标检测技术研究综述[J]. 陆军工程大学学报, 2025, 4(1): 35-46. |
| Wang Changlong, Ji Jingyu, Zhao Yuefei, et al. Review of object detection technology for remote sensing images from Unmanned Aerial Vehicles (UAVs)[J]. Journal of Army Engineering University of PLA, 2025, 4(1): 35-46. | |
| [2] | Zhou L, Liu Z, Zhao H, et al. A multi-scale object detector based on coordinate and global information aggregation for UAV aerial images[J]. Remote Sensing, 2023, 15(14): No.3468. |
| [3] | Girshick R. Fast R-CNN[C]// ICCV 2015. Piscataway: IEEE, 2015: 1440-1448. |
| [4] | 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, 2017, 39(6): 1137-1149. |
| [5] | He K, Gkioxari G, Dollár P, et al. Mask R-CNN[C]// ICCV 2017. Piscataway: IEEE, 2017: 2980-2988. |
| [6] | Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]// CVPR 2016. Piscataway: IEEE, 2016: 779-788. |
| [7] | Redmon J, Farhadi A. YOLOv3: an incremental improvement[PP/OL]. V1. arXiv (2018-04-08) [2025-04-02].. |
| [8] | Khanam R, Hussain M. What is YOLOv5: a deep look into the internal features of the popular object detector[PP/OL]. V1. arXiv (2024-07-30) [2025-04-07].. |
| [9] | Li C, Li L, Jiang H, et al. YOLOv6: a single-stage object detection framework for industrial applications[PP/OL]. V1. arXiv (2022-09-07) [2025-04-08].. |
| [10] | Yaseen M. What is YOLOv9: an in-depth exploration of the internal features of the next-generation object detector[PP/OL]. V1. arXiv (2024-09-12) [2025-04-11].. |
| [11] | Wang A, Chen H, Liu L, et al. YOLOv10: real-time end-to-end object detection[C]// NeurIPS 2024. Red Hook: Curran Associates Inc., 2024: 107984-108011. |
| [12] | Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(2): 318-327. |
| [13] | Liu W, Anguelov D, Erhan D, et al. SSD: single shot multiBox detector[C]// ECCV 2016, LNCS 9905. Cham: Springer, 2016: 21-37. |
| [14] | Tian Z, Shen C, Chen H, et al. FCOS: a simple and strong anchor-free object detector[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(4): 1922-1933. |
| [15] | Zhao Y, Lv W, Xu S, et al. DETRS beat YOLOS on real-time object detection[C]// CVPR 2024. Piscataway: IEEE, 2024: 16965-16974. |
| [16] | Kong Y, Shang X, Jia S. Drone-DETR: efficient small object detection for remote sensing image using enhanced RT-DETR model[J]. Sensors, 2024, 24(17): No.5496. |
| [17] | Zhang H, Ma Z, Li X. RS-DETR: an improved remote sensing object detection model based on RT-DETR[J]. Applied Sciences, 2024, 14(22): No.10331. |
| [18] | Wang X, Chen H. HPS-DETR: enhancing small object detection with lightweight feature extraction and Transformer integration[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: No.5937420. |
| [19] | Wang S, Jiang H, Li Z, et al. PHSI-RTDETR: a lightweight infrared small target detection algorithm based on UAV aerial photography[J]. Drones, 2024, 8(6): No.240. |
| [20] | Shi J Y, Chen H. Multi-scale feature enhancement pyramid, enhance the detection capability for small targets[PP/OL]. V1. Research Square (2024-10-21) [2025-05-21].. |
| [21] | Sunkara R, Luo T. No more strided convolutions or pooling: a new CNN building block for low-resolution images and small objects[C]// ECML PKDD 2022, LNCS 13715. Cham: Springer, 2022: 443-459. |
| [22] | Wang C Y, Mark Liao H Y, Wu Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]// CVPRW 2020. Piscataway: IEEE, 2020: 1571-1580. |
| [23] | Cui Y, Ren W, Knoll A. Omni-kernel network for image restoration[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(2): 1426-1434. |
| [24] | Chen Z, He Z, Lu Z M. DEA-Net: single image dehazing based on detail-enhanced convolution and content-guided attention[J]. IEEE Transactions on Image Processing, 2024, 33: 1002-1015. |
| [25] | Chen J R, Kao S H, He H, et al. Run, don’t walk: chasing higher FLOPS for faster neural networks[C]// CVPR 2023. Piscataway: IEEE, 2023: 12021-12031. |
| [26] | Xiao Z, Liu Q, Tang G, et al. Elliptic Fourier transformation-based histograms of oriented gradients for rotationally invariant object detection in remote-sensing images[J]. International Journal of Remote Sensing, 2015, 36(2): 618-644. |
| [27] | Long Y, Gong Y, Xiao Z, et al. Accurate object localization in remote sensing images based on convolutional neural networks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2017, 55(5): 2486-2498. |
| [28] | Suo J, Wang T, Zhang X, et al. HIT-UAV: a high-altitude infrared thermal dataset for Unmanned Aerial Vehicle-based object detection[J]. Scientific Data, 2023, 10: No.227. |
| [29] | Wei H, Liu X, Xu S, et al. DWRSeg: rethinking efficient acquisition of multi-scale contextual information for real-time semantic segmentation[PP/OL]. V3. arXiv (2023-09-13) [2025-07-02].. |
| [30] | Xia Z, Pan X, Song S, et al. Vision Transformer with deformable attention[C]// CVPR 2022. Piscataway: IEEE, 2022: 4784-4793. |
| [31] | Song Y, Zhou Y, Qian H, et al. Rethinking performance gains in image dehazing networks[PP/OL]. V1. arXiv (2022-09-23) [2025-06-23].. |
| [32] | Liu Z, Lin Y, Cao Y, et al. Swin Transformer: hierarchical Vision Transformer using shifted windows[C]// ICCV 2021. Piscataway: IEEE, 2021: 9992-10002. |
| [33] | 苏佳,杨梦凡,张柏杨,等. 改进RT-DETR的无人机小目标检测算法[J]. 微电子学与计算机, 2025, 42(7): 114-125. |
| Su Jia, Yang Mengfan, Zhang Boyang, et al. Enhancing RT-DETR for small object detection in UAV[J]. Microelectronics and Computer, 2025, 42(7): 114-125. |
| [1] | 汤莉, 张健宇, 姚睿. 基于改进YOLOv11的雾天目标检测算法[J]. 《计算机应用》唯一官方网站, 2026, 46(8): 2630-2639. |
| [2] | 张纾豪, 何坤金, 徐佳晨, 沙河山, 陈正鸣. 融合透视校正与轻量注意力机制的轮毂缺陷检测方法[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 2007-2015. |
| [3] | 朱松浩, 赵芝芸, 王梦灵. 基于轻量残差与亮度感知动态特征融合的低光图像增强网络[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1936-1946. |
| [4] | 吕超, 马歌谣. 基于冗余特征抑制的轻量级人体姿态估计网络[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1973-1980. |
| [5] | 尹秋燕, 丁婧, 聂志刚. 无人机航拍视角下的人体姿态估计算法YOLO-AirPose[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1989-1997. |
| [6] | 杜艺, 续明进, 孔佳仪, 王力瑶, 赵晨. 基于YOLOv11的低秩自适应参数高效微调算法[J]. 《计算机应用》唯一官方网站, 2026, 46(6): 1738-1745. |
| [7] | 吴闵奇, 杨元华, 李航, 胡雅琴, 汤智豪, 梅腾. 基于图Transformer和RT-DETR的轻量化水下小目标检测[J]. 《计算机应用》唯一官方网站, 2026, 46(5): 1586-1595. |
| [8] | 邢长征, 郑鑫, 贾迪, 梁浚锋. 基于自适应注意力与嵌套感受野改进DeepLabV3+方法[J]. 《计算机应用》唯一官方网站, 2026, 46(5): 1408-1415. |
| [9] | 张红瑞, 冯威铭, 杨潞霞, 马永杰. 基于YOLO11改进的水下小目标检测算法CSAF-YOLO[J]. 《计算机应用》唯一官方网站, 2026, 46(5): 1578-1585. |
| [10] | 严心怡, 朱灵龙, 张永宏. 面向复杂交通场景的多尺度实时人车检测方法CDC-DETR[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1283-1291. |
| [11] | 何帅, 邓春华. 基于YOLO-World的少样本学习目标检测算法[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1275-1282. |
| [12] | 张永兵, 闫丽蓉, 唐晓芬. 渐进式双阶段模态交互的单域泛化目标检测[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1264-1274. |
| [13] | 陈鹏, 李旭, 余肖生. RGB-D双流镜像伪装目标检测网络[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1253-1263. |
| [14] | 郭阳, 王海亮, 高需, 王海涛, 王翌博. BEV三维目标检测算法体系综述[J]. 《计算机应用》唯一官方网站, 2026, 46(4): 1238-1252. |
| [15] | 于银山, 唐旭, 丁明鉴, 黄文凯, 毕嘉文, 谭国辰. 基于YOLOv10的实时车辆检测算法[J]. 《计算机应用》唯一官方网站, 2026, 46(3): 950-958. |
| 阅读次数 | ||||||
|
全文 |
|
|||||
|
摘要 |
|
|||||