基于多阶段蒸馏的无人机图像时敏目标增量检测算法

doi:10.7527/S1000-6893.2025.31959

电子电气工程与控制

本期目录 | 过刊浏览 | 高级检索

前一篇 |

基于多阶段蒸馏的无人机图像时敏目标增量检测算法

成桢灏, 杨小冈(), 卢瑞涛, 张涛, 王思宇

火箭军工程大学导弹工程学院，西安 710025

收稿日期:2025-03-10 修回日期:2025-05-23 接受日期:2025-06-23 出版日期:2025-07-04 发布日期:2025-07-03
通讯作者: 杨小冈 E-mail:doctoryxg@163.com
基金资助:
陕西省重点研发计划(2024CY2-GJHX-42);国家自然科学基金(62276274);陕西省三秦英才特殊支持计划(2024-SQ-001)

Multi-stage distillation for incremental detection of time-sensitive targets in UAV images

Zhenhao CHENG, Xiaogang YANG(), Ruitao LU, Tao ZHANG, Siyu WANG

College of Missile Engineering，Rocket Force University of Engineering，Xi’an 710025，China

Received:2025-03-10 Revised:2025-05-23 Accepted:2025-06-23 Online:2025-07-04 Published:2025-07-03
Contact: Xiaogang YANG E-mail:doctoryxg@163.com
Supported by:
Key Research and Development Program of Shaanxi Province(2024CY2-GJHX-42);National Natural Science Foundation of China(62276274);Shaanxi SanqinElite Special Support Program(2024-SQ-001)

摘要/Abstract

摘要：

针对当前无人机图像时敏目标类增量检测面临的灾难性遗忘、过拟合以及难以适配密集检测器特性导致检测精度受限等问题，提出了一种基于多阶段蒸馏的时敏目标增量检测算法，算法主要包含基于连续Wasserstein距离的类间蒸馏（WICD）模块、基于原型引导的类内一致性蒸馏（PGICD）模块以及交叉预测自适应蒸馏（CAD）模块。WICD模块从特征图和语义查询向量中捕捉类间特征差异，利用高斯分布与连续Wasserstein距离，增强类间区分性。PGICD模块通过最小化教师网络和学生网络中实例的高层语义查询和低层特征图的原型差异，实现类内特征有效传递，增强类内一致性。CAD模块通过动态调整分类和回归分支的蒸馏权重，优化交叉预测蒸馏过程，缓解了增量学习中灾难性遗忘问题，提升了模型在复杂场景下的检测精度。在SIMD和MAR20数据集上的实验结果显示，所提方法在各类型的单步和多步增量场景下均表现优异，平均精度（AP）相比传统方法有显著提升。在SIMD数据集8类+7类的增量场景下，AP高达70.8%，与上限绝对差距为1.7%，相对差距为2.3%。在MAR20数据集10类+10类的增量场景下，AP高达60.2%，与上限的绝对差距为2.3%，相对差距为3.6%。此外，通过消融实验验证了各模块有效性，所得结果表明各模块有效地提升了无人机图像时敏目标增量检测性能。

关键词: 增量目标检测, 知识蒸馏, 无人机图像, 时敏目标检测, 交叉预测蒸馏

Abstract:

To address the problems of catastrophic forgetting， overfitting， and difficulty in adapting dense detector characteristics leading to limited detection accuracy faced by the current time-sensitive target class incremental object detection of UAV images， this paper proposes a time-sensitive target incremental detection algorithm based on multi-stage distillation. The proposed method includes three key modules： the continuous Wasserstein distance-based Inter-Class Distillation （WICD） module， the Prototype-Guided Intra-class Consistency Distillation （PGICD） module， and the Cross-prediction Adaptive Distillation （CAD） module. The WICD module captures the inter-class feature differences from the feature graphs and semantic query vectors using Gaussian distributions with continuous Wasserstein distances to enhance the inter-class discriminability. The PGICD module enhances the inter-class discriminative properties by minimizing the high-level semantic query of the instances in both the teacher’s network and student’s network， and the low-level feature graphs with the prototype differences to achieve effective intra-class feature transfer and enhance intra-class consistency. The CAD module optimizes the cross-prediction distillation process by dynamically adjusting the distillation weights of the classification and regression branches， mitigating the problem of catastrophic forgetting in incremental learning， and improving the model’s detection accuracy in complex scenarios. Experimental results on SIMD and MAR20 datasets show that the proposed method performs well in all types of one-step and multi-step incremental scenarios， and the Average Precision （AP） is significantly improved compared with traditional methods. The AP is as high as 70.8% in the incremental scenario of SIMD dataset with 8+7 classes， with an absolute gap of 1.7% and a relative gap of 2.3% from the upper limit. The results also show that the AP is as high as 60.2% in the incremental scenario of MAR20 dataset with 10+10 classes， with an absolute gap of 2.3% and a relative gap of 3.6% from the upper limit. In addition， the effectiveness of each module is verified by ablation experiments， which effectively improves the incremental detection performance of time-sensitive targets in UAV images.

Key words: incremental object detection, knowledge distillation, UAV images, time-sensitive target detection, cross-prediction distillation

中图分类号:

成桢灏, 杨小冈, 卢瑞涛, 张涛, 王思宇. 基于多阶段蒸馏的无人机图像时敏目标增量检测算法[J]. 航空学报, 2025, 46(24): 331959.

Zhenhao CHENG, Xiaogang YANG, Ruitao LU, Tao ZHANG, Siyu WANG. Multi-stage distillation for incremental detection of time-sensitive targets in UAV images[J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(24): 331959.

图/表 14

图 1

表 1

表 2

表 3

表 4

表 5

图 2

图 3

表 6

表 7

表 8

表 9

图 4

图 5

参考文献 34

[1]	ZHANG X R， ZHANG T Y， WANG G C， et al. Remote sensing object detection meets deep learning： A metareview of challenges and advances［J］. IEEE Geoscience and Remote Sensing Magazine， 2023， 11（4）： 8-44.
[2]	赵其昌，吴一全，苑玉彬. 光学遥感图像舰船目标检测与识别方法研究进展［J］. 航空学报， 2024， 45（8）： 029025.
	ZHAO Q C， WU Y Q， YUAN Y B. Progress of ship detection and recognition methods in optical remote sensing images［J］. Acta Aeronautica et Astronautica Sinica， 2024， 45（8）： 029025 （in Chinese）.
[3]	刘延芳，佘佳宇，袁秋帆，等. 无人机遥感图像实时小目标检测方法［J］. 航空学报， 2024， 45（14）： 630119.
	LIU Y F， SHE J Y， YUAN Q F， et al. Real-time small target detection networks for UAV remote sensing［J］. Acta Aeronautica et Astronautica Sinica， 2024， 45（14）： 630119 （in Chinese）.
[4]	肖欣林，施伟超，郑向涛，等. 基于多模型协同的舰船目标检测［J］. 航空学报， 2024， 45（14）： 630241.
	XIAO X L， SHI W C， ZHENG X T， et al. Multiple models collaboration for ship detection［J］. Acta Aeronautica et Astronautica Sinica， 2024， 45（14）： 630241 （in Chinese）.
[5]	LI Z， WANG Y C， ZHANG N， et al. Deep learning-based object detection techniques for remote sensing images： A survey［J］. Remote Sensing， 2022， 14（10）： 2385.
[6]	ZHOU D W， WANG Q W， QI Z H， et al. Class-incremental learning： A survey［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2024， 46（12）： 9851-9873.
[7]	MASANA M， LIU X L， TWARDOWSKI B， et al. Class-incremental learning： Survey and performance evaluation on image classification［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2023， 45（5）： 5513-5533.
[8]	MA B T， CONG Y， REN Y. IOSL： Incremental open set learning［J］. IEEE Transactions on Circuits and Systems for Video Technology， 2024， 34（4）： 2235-2248.
[9]	KANG M X， ZHANG J P， ZHANG J M， et al. Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation［C］∥ 2023 IEEE/CVF International Conference on Computer Vision （ICCV）. Piscataway： IEEE Press， 2023： 18848-18858.
[10]	方维维，陈爱方，孟娜，等. 基于知识蒸馏的目标检测模型增量深度学习方法［J］. 工程科学与技术， 2022， 54（6）： 59-66.
	FANG W W， CHEN A F， MENG N， et al. Incremental deep learning method for object detection model based on knowledge distillation［J］. Advanced Engineering Sciences， 2022， 54（6）： 59-66 （in Chinese）.
[11]	KIM J， CHO H， KIM J， et al. SDDGR： Stable diffusion-based deep generative replay for class incremental object detection［C］∥ 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2024： 28772-28781.
[12]	KIM J， KU Y， KIM J， et al. VLM-PL： Advanced pseudo labeling approach for class incremental object detection via vision-language model［C］∥ 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Piscataway： IEEE Press， 2024： 4170-4181.
[13]	PENG C， ZHAO K， LOVELL B C. Faster ILOD： Incremental learning for object detectors based on faster RCNN［J］. Pattern Recognition Letters， 2020， 140： 109-115.
[14]	FENG T， WANG M， YUAN H J. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation［C］∥ 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2022： 9417-9426.
[15]	LIU Y Y， CONG Y， GOSWAMI D， et al. Augmented box replay： Overcoming foreground shift for incremental object detection［C］∥ 2023 IEEE/CVF International Conference on Computer Vision （ICCV）. Piscataway： IEEE Press， 2023： 11333-11343.
[16]	LIU Y Y， SCHIELE B， VEDALDI A， et al. Continual detection transformer for incremental object detection［C］∥ 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2023： 23799-23808.
[17]	LI W Z， ZHOU J W， LI X， et al. InfRS： Incremental few-shot object detection in remote sensing images［J］. IEEE Transactions on Geoscience and Remote Sensing， 2024， 62： 5644314.
[18]	张涛，杨小冈，卢孝强，等. Dense RFB和LSTM遥感图像舰船目标检测［J］. 遥感学报， 2022， 26（9）： 1859-1871.
	ZHANG T， YANG X G， LU X Q， et al. Ship detection in remote sensing image based on dense RFB and LSTM［J］. National Remote Sensing Bulletin， 2022， 26（9）： 1859-1871 （in Chinese）.
[19]	JOSEPH K J， RAJASEGARAN J， KHAN S， et al. Incremental object detection via meta-learning［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2022， 44（12）： 9209-9216.
[20]	BAI L， SONG H， FENG T， et al. Revisiting class-incremental object detection： An efficient approach via intrinsic characteristics alignment and task decoupling［J］. Expert Systems with Applications， 2024， 257： 125057.
[21]	MO Q J， GAO Y P， FU S H， et al. Bridge past and future： Overcoming information asymmetry in incremental object detection［C］∥ Computer Vision-ECCV 2024. Cham： Springer Nature Switzerland， 2024： 463-480.
[22]	LU X N， DIAO W H， LI J X， et al. Few-shot incremental object detection in aerial imagery via dual-frequency prompt［J］. IEEE Transactions on Geoscience and Remote Sensing， 2024， 62： 5624017.
[23]	YU Q Z， ZHU K， WANG W， et al. Incremental object detection with image-level labels［J］. IEEE Transactions on Artificial Intelligence， 2024， 5（5）： 2331-2341.
[24]	DONG N， ZHANG Y Q， DING M L， et al. Incremental-DETR： Incremental few-shot object detection via self-supervised learning［J］.Proceedings of the AAAI Conference on Artificial Intelligence， 2023， 37（1）： 543-551.
[25]	LI J Y， CAO Z J， GAN Q， et al. Class-incremental SAR obeject detection via adaptive distributed response distillation［C］∥ IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. Piscataway： IEEE Press， 2024： 9922-9925.
[26]	DONG N， ZHANG Y， DING M， et al. Bridging non-co-occurrence with unlabeled in-the-wild data for incremental object detection［C］∥ Proceedings of the 35th International Conference on Neural Information Processing Systems， New York：ACM， 2021：30492-30503.
[27]	LI D， TASCI S， GHOSH S， et al. RILOD： Near real-time incremental learning for object detection at the edge［C］∥ Proceedings of the 4th ACM/IEEE Symposium on Edge Computing. New York：ACM， 2019： 113-126.
[28]	CERMELLI F， GERACI A， FONTANEL D， et al. Modeling missing annotations for incremental learning in object detection［C］∥ 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Piscataway： IEEE Press， 2022： 3699-3709.
[29]	WANG Y J， CHEN L Q， ZHAO T M， et al. High-dimension prototype is a better incremental object detection learner［C］∥ The Thirteenth International Conference on Learning Representations， 2025： 1-17.
[30]	WANG J B， CHEN Y M， ZHENG Z H， et al. CrossKD： Cross-head knowledge distillation for object detection［C］∥ 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2024： 16520-16530.
[31]	HAROON M， SHAHZAD M， FRAZ M M. Multisized object detection using spaceborne optical imagery［J］. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing， 2020， 13： 3032-3046.
[32]	禹文奇，程塨，王美君，等.MAR20：遥感图像军用飞机目标识别数据集［J］.遥感学报，2023，27（12）：2688-2696.
	YU W Q， CHENG G， WANG M J， et al. MAR20： A benchmark for military aircraft recognition in remote sensing images［J］. National Remote Sensing Bulletin， 2023， 27（12）： 2688-2696 （in Chinese）.
[33]	LIU L Y， KUANG Z H， CHEN Y M， et al. IncDet： In defense of elastic weight consolidation for incremental object detection［J］. IEEE Transactions on Neural Networks and Learning Systems， 2021， 32（6）： 2306-2319.
[34]	MENEZES A G， DE MOURA G， ALVES C， et al. Continual object detection： A review of definitions， strategies， and challenges［J］. Neural Networks， 2023， 161： 476-493.

E-mail：hkxb@buaa.edu.cn

关于我们

期刊社服务

专业学科

封面文章

友情链接

主管单位：中国科学技术协会主办单位：中国航空学会北京航空航天大学

距离度量	原理	优势	局限性
欧氏距离	计算2点间直线距离	计算简单直观，低维数据能快速衡量向量差异	未考虑特征分布，高维空间中数据稀疏，难以反映真实类间差异
余弦相似度	计算向量夹角余弦值，衡量方向相似性	对数据尺度不敏感，专注方向差异，适合比较向量方向一致性	忽略向量模长，方向相同但模长差异大时易误判
KL散度	衡量2个概率分布差异，计算近似分布时损失的信息	擅长度量概率分布差异，知识蒸馏中用于学习教师模型分布	对分布假设严格，可能丢失特征不确定性和多样性，在复杂场景下的表现较差
曼哈顿距离	计算各维度绝对差值之和（标准坐标系上的轴距总和）	计算高效、对异常值不敏感，适合网格数据或降低极端值影响的场景	未考虑维度相关性，高维复杂数据中难捕捉真实差异，单一维度大差异影响显著
Wasserstein 距离	衡量两概率分布的最传输成本，通过最优传输路径计算分布差异	考虑特征分布结构，高维处理有效，对分布变化敏感，高斯建模计算高效	计算复杂度相对较高，在大规模数据上计算耗时较长

Distance metrics	AP/%	AP₅₀/%	AP₇₅/%	AP_S/%	AP_M/%	AP_L/%	Upper Bound/%	AbsGap/%	RelGap/%
欧氏距离	65.3	78.2	75.1	15.6	58.9	70.8	70.2	4.9	6.9
余弦相似度	56.4	68.3	64.9	4.3	49.1	61.5	61.9	5.5	8.8
KL散度	62.1	74.5	71.3	11.2	55.4	68.3	68.8	6.7	9.7
曼哈顿距离	58.2	71.9	68.7	8.5	52.7	65.2	64.3	6.1	9.5
Wasserstein距离	69.2	80.7	77.5	20.4	61.2	72.1	72.5	3.3	4.5

Distance metrics	AP/%	AP₅₀/%	AP₇₅/%	AP_S/%	AP_M/%	AP_L/%	Upper Bound/%	AbsGap/%	RelGap/%
欧氏距离	56.8	69.3	66.1	9.8	49.2	52.5	61.2	4.4	7.2
余弦相似度	48.1	59.7	55.4	2.3	38.5	49.1	54.6	6.5	11.9
KL散度	53.2	65.8	62.3	6.7	45.1	51.3	58.4	5.2	8.9
曼哈顿距离	50.4	62.1	58.9	4.1	41.3	52.7	55.3	4.9	8.8
Wasserstein距离	58.3	71.8	68.2	10.2	51.4	56.1	62.5	4.2	6.7

Scenarios	Methods	AP/%	AP₅₀/%	AP₇₅/%	AP_S/%	AP_M/%	AP_L/%	Upper Bound/%	AbsGap/%	RelGap/%
8类+7类	Faster ILOD	54.2	72.4	63.2	6.7	46.3	57.3	60.3	6.1	10.1
	iOD	56.8	74.1	64.4	8.2	48.5	59.7	61.1	4.3	7.1
	ERD	58.1	75.9	66.9	8.8	49.1	63.6	62.4	4.3	6.9
	ABR-IOD	57.7	74.5	64.9	8.4	48.9	61.4	61.6	3.9	6.3
	ECOD	61.2	78.3	71.9	10.3	49.4	68.1	63.8	2.6	4.1
	Efficient-IOD	66.7	79.6	77.8	19.2	59.3	70.9	69.6	2.9	4.2
	Ours	70.8	82.4	79.4	21.2	62.3	73.3	72.5	1.7	2.3
10类+5类	Faster ILOD	49.6	67.1	59.2	5.2	43.3	53.9	60.3	10.7	17.7
	iOD	52.2	68.8	60.5	7.5	45.1	56.2	61.1	8.9	14.6
	ERD	54.5	70.4	62.2	8.4	46.7	59.1	62.4	7.9	12.7
	ABR-IOD	52.9	69.7	59.9	7.7	46.4	58.4	61.6	8.7	14.1
	ECOD	60.7	77.4	72.2	10.1	49.8	67.3	63.8	3.1	4.9
	Efficient-IOD	63.7	77.6	75.6	18.3	50.6	68.1	69.6	5.9	8.5
	Ours	69.2	80.7	77.5	20.4	61.2	72.1	72.5	3.3	4.5
12类+3类	Faster ILOD	46.2	64.4	56.7	4.8	41.2	51.4	60.3	14.1	23.4
	iOD	48.3	65.1	57.6	6.7	43.5	54.7	61.1	12.8	20.9
	ERD	51.8	66.7	62.1	10.6	49.5	56.4	62.4	10.6	16.9
	ABR-IOD	50.1	66.5	57.4	7.2	44.8	57.1	61.6	11.5	18.7
	ECOD	58.3	70.2	68.3	11.3	45.3	64.5	63.8	5.5	8.6
	Efficient-IOD	60.9	75.4	67.6	11.7	45.8	64.2	69.6	8.7	12.5
	Ours	67.9	80.1	76.6	19.8	60.4	70.8	72.5	4.6	6.3

Scenarios	方法	AP/%	AP₅₀/%	AP₇₅/%	AP_S/%	AP_M/%	AP_L/%	Upper Bound/%	AbsGap/%	RelGap/%
10类+10类	Faster ILOD	40.2	55.3	52.3	3.9	38.6	44.7	48.3	8.1	16.8
	iOD	42.3	58.1	54.2	4.4	41.3	47.9	50.2	7.9	15.7
	ERD	46.8	62.3	56.8	14.1	43.7	50.5	54.3	7.5	13.8
	ABR-IOD	47.6	64.5	57.7	5.5	44.5	51.2	53.8	6.2	11.5
	ECOD	48.9	67.2	60.8	4.8	45.8	53.4	54.5	5.6	10.3
	Efficient-IOD	53.7	70.8	67.4	10.6	50.9	52.3	57.2	3.5	6.1
	Ours	60.2	73.9	70.3	12.7	53.3	56.8	62.5	2.3	3.6
12类+8类	Faster ILOD	38.8	53.3	51.2	3.3	37.2	42.7	48.3	9.5	19.7
	iOD	40.9	56.4	53.4	4.1	40.1	45.9	50.2	9.3	18.5
	ERD	42.2	59.8	56.3	5.2	42.3	48.9	54.3	12.1	22.2
	ABR-IOD	44.7	61.5	53.7	3.5	41.5	49.1	53.8	9.1	17.1
	ECOD	47.5	62.8	58.2	2.7	44.8	51.5	54.5	7.0	12.8
	Efficient-IOD	50.7	69.4	63.4	5.6	46.6	55.9	57.2	6.5	11.4
	Ours	58.3	71.8	68.2	10.2	51.4	56.1	62.5	4.2	6.7
15类+5类	Faster ILOD	36.4	52.1	49.2	2.6	36.3	41.9	48.3	11.9	24.6
	iOD	39.3	55.7	52.4	3.8	39.4	45.1	50.2	10.9	21.7
	ERD	40.7	56.4	54.8	4.9	41.7	48.3	54.3	13.6	33.4
	ABR-IOD	41.1	57.5	49.2	4.1	40.2	48.7	53.8	12.7	23.6
	ECOD	46.3	62.2	56.1	2.8	44.6	52.5	54.5	8.2	15.1
	Efficient-IOD	49.3	65.9	60.4	4.6	42.9	53.7	57.2	7.9	13.8
	Ours	57.7	70.6	67.5	9.8	50.2	55.3	62.5	4.8	7.6

基于多阶段蒸馏的无人机图像时敏目标增量检测算法

Multi-stage distillation for incremental detection of time-sensitive targets in UAV images

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 14

参考文献 34

相关文章 2

编辑推荐

Metrics

本文评价

WICD	PGICD	CAD	Step1				Step2				Step3
WICD	PGICD	CAD	AP/%	AP₅₀/%	AP₇₅/%	RelGap/%	AP/%	AP₅₀/%	AP₇₅/%	RelGap/%	AP/%	AP₅₀/%	AP₇₅/%	RelGap/%
			61.3	71.2	69.6	15.4	58.4	69.4	67.3	22.5	56.2	66.7	64.5	26.1
✓			66.1	75.5	73.8	8.8	63.6	74.1	71.8	12.2	61.6	71.3	69.8	15.0
✓	✓		69.2	79.8	75.6	4.6	66.5	75.9	72.3	8.3	64.8	74.2	71.1	10.6
✓		✓	69.5	80.6	77.3	4.1	66.9	76.6	75.5	7.7	65.3	75.2	73.5	9.9
	✓	✓	70.2	82.1	78.3	3.1	67.8	79.2	76.1	6.5	67.1	77.2	74.1	7.4
✓	✓	✓	71.8	84.9	79.6	1.0	69.7	81.3	77.9	3.9	68.9	80.2	76.5	4.9

[1]	丛润民, 孙豪言, 罗宇轩, 方豪. 基于类关系挖掘的遥感图像广义小样本分割方法[J]. 航空学报, 2025, 46(23): 631694-631694.
[2]	冒国韬, 邓天民, 于楠晶. 基于多尺度分割注意力的无人机航拍图像目标检测算法[J]. 航空学报, 2023, 44(5): 326738-326738.