基于扩散超分重建的小目标检测方法

doi:10.12305/j.issn.1001-506X.2026.09.06

摘要/Abstract

摘要：

随着无人机遥感技术的广泛应用，小目标检测技术受到越来越多的关注。然而，由于图像分辨率受限、噪声干扰以及复杂背景等问题，现有检测方法的精度和鲁棒性依然面临挑战。针对以上问题，提出基于扩散结构的超分重建（super-resolution reconstruction with diffusion structure，SRRDS)的小目标检测方法。首先，针对图像中常见的噪声干扰等问题，提出一种扩散模型驱动退化数据生成模块，通过模拟实际成像系统的物理退化过程，结合双阶段物理约束退化引擎与可控噪声增强模块，生成真实物理退化特征的低质图像，提升训练数据与实际场景的匹配性。其次，针对低分辨率图像纹理细节缺失和边缘信息退化问题，提出多尺度残差生成器模块。通过引入多尺度分支结构、像素重组操作和残差生成块，有效增强输入图像的层次信息表征能力。同时，利用生成对抗网络优化图像质量，实现高精度的超分辨率重建。在AI-TOD数据集和VisDrone-2019数据集中平均精度（average precision，AP）达到了14.1%和25.7%，AP50分别达到了32.9%和41.4%。实验结果表明，与常规检测模型及数据增强方法对比，扩散噪声的引入能够模拟真实场景退化现象，且提高图像分辨率，减少小尺度目标的漏检和虚警现象。

关键词: 扩散结构, 数据增强, 超分辨率重建, 目标检测

Abstract:

As unmanned aerial remote sensing technology becomes increasingly widespread, small target detection technology is receiving growing attention. However, due to limitations in image resolution, noise interference, and complex backgrounds, the accuracy and robustness of existing detection methods still face significant challenges. Therefore, a small target detection method is proposed based on super-resolution reconstruction with diffusion structure （SRRDS）. Firstly, to address common issues such as noise interference in images, a diffusion-model-driven degraded data generation module is proposed. By simulating the physical degradation process of actual imaging systems and integrating a dual-stage physically constrained degradation engine with a controllable noise enhancement module, this module generates low-quality images with realistic physical degradation characteristics, improving the alignment between training data and real-world scenarios. Secondly, for issues such as the loss of texture details and edge information in low-resolution images, a multi-scale residual generator module is proposed. By incorporating a multi-scale branch structure, pixel rearrangement operations, and residual generation blocks, this module effectively enhances input image hierarchical information representation. Additionally, it utilizes generative adversarial network to optimize image quality, achieving high-precision super-resolution reconstruction. On the AI-TOD dataset and the VisDrone-2019 dataset, average precision （AP） achieved 14.1% and 25.7%, while AP50 reached 32.9% and 41.4%, respectively. Experimental results demonstrate that compared to conventional detection models and data enhancement methods, the introduction of diffusion noise can simulate real-world scenarios degradation phenomena, improve image resolution, and reduce missed detections and false alarms for small-scale targets.

Key words: diffusion structure, data augmentation, super-resolution reconstruction, target detection

中图分类号:

TP 391.41

宋润泽, 崔洪博, 赵子龙, 刘建飞. 基于扩散超分重建的小目标检测方法[J]. 系统工程与电子技术, 2026, 48(9): 2931-2938.

Runze Song, Hongbo Cui, Zilong Zhao, Jianfei Liu. Small target detection method based on diffusion super-resolution reconstruction[J]. Systems Engineering and Electronics, 2026, 48(9): 2931-2938.

图/表 12

图1

图2

图3

图4

图5

图6

表1

表2

图7

表3

图8

表4

参考文献 30

1	刘颖, 刘红燕, 范九伦, 等. 基于深度学习的小目标检测研究与应用综述[J]. 电子学报, 2020, 48 (3): 590. doi: 10.3969/j.issn.0372-2112.2020.03.024
2	吕璐, 程虎, 朱鸿泰, 等. 基于深度学习的目标检测研究与应用综述[J]. 电子与封装, 2022, 22 (1): 72. doi: 10.16257/j.cnki.1681-1070.2022.0114
3	李波, 黄晶益, 万开方, 等. 基于深度强化学习的无人机系统应用研究综述[J]. 战术导弹技术, 2023 (1): 58. doi: 10.16358/j.issn.1009-1300.20220211
4	Wang J W, Yang W, Guo H W, et al. Tiny object detection in aerial images[C]// 25th International Conference on Pattern Recognition, 2021: 3791.
5	汤文兵, 翁正阳. 基于WCCE-YOLO的无人机遥感图像小目标检测[EB/OL]. [2025-06-20]. https://link.cnki.net/urlid/50.1155.n.20250818.1531.004.
6	李婧. 基于改进YOLOv12的无人机航拍小目标检测研究[J]. 信息记录材料, 2025, 26 (8): 235.
7	魏松林, 林静敏, 蔡艺军. 基于改进YOLOv8的无人机海上搜救目标检测算法[J]. 绵阳师范学院学报, 2025, 44 (8): 36. doi: 10.16276/j.cnki.cn51-1670/g.2025.08.005
8	赵浩权. 面向无人机图像的小目标检测方法研究[D]. 长春: 吉林大学, 2023.
9	Yang X, Yan J C, Ming Q, et al. Rethinking rotated object detection with Gaussian wasserstein distance loss[C]// International Conference on Machine Learning, 2021: 11830.
10	沈焕锋, 李平湘, 张良培, 等. 图像超分辨率重建技术与方法综述[J]. 光学技术, 2009, 35 (2): 7. doi: 10.3321/j.issn:1002-1582.2009.02.032
11	Ledig C, Theis L, Huszár F, et al. Photo-realistic single image super-resolution using a generative adversarial network[C]// IEEE Conference on Computer Vision and Pattern Recognition, 2017: 4681.
12	San-Roman R, Nachmani E, Wolf L. Noise estimation for generative diffusion models[EB/OL]. [2025-04-06]. https://arxiv.org/abs/2104.02600.
13	Liu Z L, Xu G D, Yang H H, et al. Suppress-and-refine framework for end-to-end 3D object detection[EB/OL]. [2025-03-18]. https://arxiv.org/abs/2103.10042.
14	Zhang J Q, Lei J, Xie W Y, et al. SuperYOLO: super resolution assisted object detection in multimodal remote sensing imagery[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61, 5605415. doi: 10.52783/jes.3664
15	邵凯, 李浩刚, 梁燕, 等. 跨尺度特征融合的遥感微小目标检测算法[J]. 系统工程与电子技术, 2025, 47 (5): 1421. doi: 10.12305/j.issn.1001-506X.2025.05.05
16	成科扬, 荣兰, 蒋森林, 等. 面向目标检测的双驱自适应遥感图像超分重建方法[J]. 北京航空航天大学学报, 2021, 48 (8): 1343. doi: 10.13700/j.bh.1001-5965.2021.0517
17	Croitoru F A, Hondru V, Ionescu R T, et al. Diffusion models in vision: a survey[J]. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2023, 45 (9): 10850. doi: 10.1109/TPAMI.2023.3261988
18	Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models[J]. Advances in Neural Information Processing Systems, 2020, 33, 6840.
19	Tan M, Pang R, Le Q V. Efficientdet: scalable and efficient object detection[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 10781.
20	曾拥华, 马均瑶, 黄朝燕, 等. 联合纯四元数与字典学习的彩色图像去噪方法[J]. 系统工程与电子技术, 2023, 45 (2): 373. doi: 10.12305/j.issn.1001-506X.2023.02.06
21	Wang J Y, Lu Y J, Wang M K, et al. An enhanced super-resolution generative adversarial network (ESRGAN) for improving image quality on construction vehicle license plates[C]// International Symposium on Advancement of Construction Management and Real Estate, 2024.
22	Schonfeld E, Schiele B, Khoreva A. A U-Net based discriminator for generative adversarial networks[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 8207.
23	Bilecen B B, Ayazoglu M. Bicubic++: slim, slimmer, slimmest-designing an industry-grade super-resolution network[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023: 1623.
24	Huang Y, Li S, Wang L, et al. Unfolding the alternating optimization for blind super resolution[J]. Advances in Neural Information Processing Systems, 2020, 33, 5632.
25	Zhang K, Liang J, Van G L, et al. Designing a practical degradation model for deep blind image super-resolution[C]// IEEE/CVF International Conference on Computer Vision, 2021: 4791.
26	Kuriakose B M. EDSR: empowering super-resolution algorithms with high-quality DIV2K images[J]. Intelligent Decision Technologies, 2023, 17 (4): 1249. doi: 10.3233/IDT-230218
27	Agustsson E, Timofte R. Ntire 2017 challenge on single image super-resolution: dataset and study[C]// IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017: 126.
28	Shermeyer J, Van Etten A. The effects of super-resolution on object detection performance in satellite imagery[C]// IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019.
29	Du D W, Zhu P F, Wen L Y, et al. VisDrone-DET2019: the vision meets drone object detection in image challenge results[C]// IEEE/CVF International Conference on Computer Vision Workshops, 2019.
30	宋存利, 柴伟琴, 张雪松. 基于改进YOLOv5算法的道路小目标检测[J]. 系统工程与电子技术, 2024, 46 (10): 3271.

方法	Set5		Set14		BSD100		Urban100		Manga109
方法	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM
Bicubic	28.42	0.8104	26.00	0.7027	25.96	0.6675	23.14	0.6577	24.89	0.7866
SRCNN	30.48	0.8628	27.50	0.7513	26.90	0.7101	24.52	0.7221	27.58	0.8555
EDSRN	32.46	0.8968	28.80	0.7876	27.71	0.7420	26.64	0.8033	31.02	0.9148
ESRGAN	32.60	0.9002	28.88	0.7896	27.76	0.7432	26.73	0.8072	31.16	0.9164
DAN	32.68	0.9008	28.92	0.7907	27.79	0.7443	26.89	0.8106	31.43	0.9172
BSRGAN	32.73	0.9011	28.99	0.7917	27.85	0.7455	27.03	0.8153	31.66	0.9196
SRRDS	32.80	0.9021	28.98	0.7909	27.91	0.7485	27.21	0.8177	31.71	0.9198

评估指标	原始图像	Bicubic	ESRGAN	BSRGAN	SRRDS
AP50	18.7	29.8	31	31.3	32.9
AP	7.44	12.3	13.1	13.2	14.1

评估指标	原始图像	Bicubic	ESRGAN	BSRGAN	SRRDS
AP50	25.9	35.7	37.7	40.6	41.4
AP	12.8	20.4	21.5	23.8	25.7

方法	扩散模型驱动退化数据生成模块	多尺度残差生成器	PSNR/dB	SSIM
基础	−	−	29.43	0.8313
	√	−	29.60	0.8314
	−	√	29.68	0.8349
SRRDS	√	√	29.73	0.8358

[1]	姜娜, 刘霖, 陈世龙. 基于证据理论的多源图像目标检测方法[J]. 系统工程与电子技术, 2026, 48(7): 2133-2143.
[2]	张霄嵩, 王鼎衡, 李昊飞, 王一朋, 张岳, 刘保荣, 贾世诚. 面向战场复杂场景的增强识别方法[J]. 系统工程与电子技术, 2026, 48(7): 2172-2183.
[3]	付卫红, 彭文洪, 刘乃安. 面向高效检测的YOLOv11轻量化SAR图像船舶检测模型[J]. 系统工程与电子技术, 2026, 48(7): 2241-2250.
[4]	赵巍山, 尤思洋, 黄丽佳, 周光尧. 基于散射拓扑增强的双流SAR飞机检测网络[J]. 系统工程与电子技术, 2026, 48(6): 1848-1858.
[5]	赵俊博, 毛宏霞, 赵慧洁, 张佑堃, 刘畅. QR分解-凸集投影协同超分辨率重建[J]. 系统工程与电子技术, 2026, 48(5): 1474-1480.
[6]	张杰, 常天庆, 王晓卫, 郝文龙, 汤鑫. 基于可见光与红外特征融合的轻量化目标检测方法[J]. 系统工程与电子技术, 2026, 48(5): 1481-1491.
[7]	王新, 周生华, 张新勋, 南静怡, 孙世坤. 全极化MIMO雷达主瓣压制干扰抑制方法[J]. 系统工程与电子技术, 2026, 48(5): 1539-1550.
[8]	刘硕, 周宇, 张哲昊, 尚嵩, 程松. 时域多脉冲相关与聚类分析的抗同频干扰技术[J]. 系统工程与电子技术, 2026, 48(4): 1186-1194.
[9]	鲁明雨, 孟飞, 叶春茂, 李璋峰, 赵庆媛. 一种先验知识辅助的扩展目标检测方法[J]. 系统工程与电子技术, 2026, 48(4): 1209-1217.
[10]	程鲲, 雷洪涛, 吕志轩. 改进YOLOv8的轻量级无人机跟踪方法[J]. 系统工程与电子技术, 2026, 48(3): 737-750.
[11]	王威, 杨勇, 韩静雯. 基于相关性的雷达海面弱目标检测技术[J]. 系统工程与电子技术, 2026, 48(3): 787-794.
[12]	缪燕子, 赵志非, 吴巍. 基于代价敏感学习DBN-XGBoost的海面小目标检测方法研究[J]. 系统工程与电子技术, 2026, 48(2): 402-409.
[13]	陈凯, 赵永波, 刘仍莉, 邓海涛, 孙龙. 机载长合成孔径时间海面运动舰船高分辨SAR成像算法[J]. 系统工程与电子技术, 2026, 48(2): 456-465.
[14]	丁昊, 韦继丰, 董云龙, 曹政, 于恒力. 高海况下应用先验信息的海上小目标检测方法[J]. 系统工程与电子技术, 2026, 48(1): 44-55.
[15]	蓝舒尧, 李宇, 张春华, 迟骋, 陈春辉. 垂直阵分集束控影区探测信号发射策略[J]. 系统工程与电子技术, 2025, 47(8): 2454-2462.