小波时频局部化无人机目标检测模型压缩研究

doi:10.7527/S1000-6893.2025.31952

Abstract

Abstract:

Current deep learning-based remote sensing object detection models rely on the powerful computing and storage resources of ground servers， and now capable of efficiently processing massive remote sensing data. However， in deployment scenarios for mobile edge devices such as Unmanned Aerial Vehicles （UAVs）， the limited computational resources， storage capacity， and energy constraints pose significant challenges for effectively deploying large-scale models. Model compression techniques， such as lightweight design and model quantization， have become critical forfacilitating the practical application of remote sensing object detection algorithms. This paper proposes a wavelet time-frequency localization-based model compression framework， using the single-stage object detection network as the baseline： By integrating depthwise separable convolution with the spatial-frequency localization characteristics of discrete wavelet transforms， a Wavelet Depthwise Separable Convolution （W-DSConv） module is constructed to achieve lightweight model reconstruction while expanding its receptive field； Leveraging wavelet frequency domain decomposition， a Wavelet Frequency-Division Quantization （W-FDQ） method for quantization-aware training is proposed， enabling independent quantization of features across different frequency bands to further compress the lightweight model. Experiments are conducted using YOLO-series models on the VisDrone2021 UAV remote sensing dataset.Results demonstrate that： The W-DSConv module reduces model parameters by 45.5% and computational load by 37.4%， while limiting detection accuracy fluctuations to within 2.2%； When applying 6-bit and 4-bit quantization via W-FDQ， the quantized models retain 95.8% and 92.9% of the floating-point model’s detection performance， respectively. This research provides novel technical insights for lightweight deployment of remote sensing object detection models on mobile platforms.

Key words: object detection, Unmanned Aerial Vehicles (UAVs), model compression, lightweight, model quantization, wavelet transforms

CLC Number:

V279

Wei HUANG, Jiahao PAN, Chu HE. Wavelet time-frequency localization-based model compression for UAV object detection[J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(23): 631952.

Figures/Tables 16

Fig.1

Fig.2

Table 1

Comparison of convolutional module

模块	参数量	计算量
传统卷积	$2 K 2 C 2$	$2 K 2 C 2 H W$
深度可分离卷积	$K 2 C + 2 C 2$	$(K 2 C + 2 C 2) H W$
MV2Block	$6 K 2 C + 18 C 2$	$(6 K 2 C + 18 C 2) H W$
SV2Block	$K 2 C + 2 C 2$	$(K 2 C + 2 C 2) H W$
GhostConv	$K 2 C + C 2$	$(K 2 C + C 2) H W$
WaveGhostConv	$K 2 C + C 2$	$K 2 C + N C + C 2 + 163 1 - 1 4 M C H W$

Table 1

Fig.3

Fig.4

Fig.5

Fig.6

Table 2

Table 3

Table 4

Table 5

Table 6

Table 7

Fig.7

Fig.8

Table 8

References 33

[1]	江波，屈若锟，李彦冬，等. 基于深度学习的无人机航拍目标检测研究综述［J］. 航空学报， 2021， 42（4）： 524519.
	JIANG B， QU R K， LI Y D， et al. Object detection in UAV imagery based on deep learning： Review［J］. Acta Aeronautica et Astronautica Sinica， 2021， 42（4）： 524519 （in Chinese）.
[2]	欧阳权，张怡，马延，等. 基于深度学习的无人机航拍目标检测与跟踪方法综述［J］. 电光与控制， 2024， 31（3）： 1-7.
	OUYANG Q， ZHANG Y， MA Y， et al. A review of UAV aerial photography target detection and tracking methods based on deep learning［J］. Electronics Optics & Control， 2024， 31（3）： 1-7 （in Chinese）.
[3]	赵禄达，胡以华，赵楠翔，等. LiDAR点云深度学习模型的压缩和部署加速方法研究现状与展望（特邀）［J］. 激光与光电子学进展， 2024， 61（20）： 2011005.
	ZHAO L D， HU Y H， ZHAO N X， et al. Review of model compression and accelerated development for deep learning in LiDAR point cloud processing （Invited）［J］. Laser & Optoelectronics Progress， 2024， 61（20）： 2011005 （in Chinese）.
[4]	CHEN F H， LI S L， HAN J L， et al. Review of lightweight deep convolutional neural networks［J］. Archives of Computational Methods in Engineering， 2024， 31（4）： 1915-1937.
[5]	王军，冯孙铖，程勇. 深度学习的轻量化神经网络结构研究综述［J］. 计算机工程， 2021， 47（8）： 1-13.
	WANG J， FENG S C， CHENG Y. Survey of research on lightweight neural network structures for deep learning［J］. Computer Engineering， 2021， 47（8）： 1-13 （in Chinese）.
[6]	SIFRE L， MALLAT S. Rigid-motion scattering for texture classification［DB/OL］. arXiv preprint： 1403.1687， 2014.
[7]	HOWARD A G， ZHU M L， CHEN B， et al. Mobilenets： Efficient convolutional neural networks for mobile vision applications［DB/OL］. arXiv preprint： 1704.04861， 2017.
[8]	SANDLER M， HOWARD A， ZHU M L， et al. MobileNetV2： Inverted residuals and linear bottlenecks［C］∥ 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE Press， 2018： 4510-4520.
[9]	ZHANG X Y， ZHOU X Y， LIN M X， et al. ShuffleNet： An extremely efficient convolutional neural network for mobile devices［C］∥ 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE Press， 2018： 6848-6856.
[10]	HAN K， WANG Y H， TIAN Q， et al. GhostNet： More features from cheap operations［C］∥ 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2020： 1577-1586.
[11]	VASU P K A， GABRIEL J， ZHU J， et al. MobileOne：An improved one millisecond mobile backbone［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2023： 7907-7917.
[12]	HAN S， MAO H Z， DALLY W J. Deep compression： Compressing deep neural networks with pruning， trained quantization and huffman coding［DB/OL］. arXiv preprint： 1510.00149， 2015.
[13]	LIU X C， YE M， ZHOU D Y， et al. Post-training quantization with multiple points： Mixed precision without mixed precision［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2021， 35（10）： 8697-8705.
[14]	NAGEL M， AMJAD R A， VAN BAALEN M， et al. Up or down？ Adaptive rounding for post-training quantization［C］∥Proceedings of the 37th International Conference on Machine Learning. New York： ACM， 2020： 7197-7206.
[15]	YUAN Z H， XUE C H， CHEN Y Q， et al. PTQ4ViT： Post-training quantization for vision transformers withtwin uniform quantization［C］∥Computer Vision-ECCV 2022. Cham： Springer， 2022： 191-207.
[16]	ESSER S K， MCKINSTRY J L， BABLANI D， et al. Learned step size quantization［DB/OL］. arXiv preprint：1902.08153， 2019.
[17]	BHALGAT Y， LEE J， NAGEL M， et al. LSQ+： Improving low-bit quantization through learnable offsets and better initialization［C］∥2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops （CVPRW）. Piscataway： IEEE Press， 2020： 2978-2985.
[18]	CHOI J， WANG Z， VENKATARAMANI S， et al. Pact： Parameterized clipping activation for quantized neural networks［DB/OL］. arXiv preprint： 1805.06085， 2018.
[19]	LIU Z C， CHENG K T， HUANG D， et al. Nonuniform-to-uniform quantization： Towards accurate quantization via generalized straight-through estimation［C］∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2022： 4932-4942.
[20]	ZHU K， HE Y Y， WU J X. Quantized feature distillation for network quantization［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2023， 37（9）： 11452-11460.
[21]	MUSA A， KAKUDI H A， HASSAN M， et al. Lightweight deep learning models for edge devices： A survey［J］. International Journal of Computer Information Systems and Industrial Management Applications， 2025， 17： 18.
[22]	杨春，张睿尧，黄泷，等. 深度神经网络模型量化方法综述［J］. 工程科学学报， 2023， 45（10）： 1613-1629.
	YANG C， ZHANG R Y， HUANG L， et al. A survey of quantization methods for deep neural networks［J］. Chinese Journal of Engineering， 2023， 45（10）： 1613-1629 （in Chinese）.
[23]	NAGEL M， FOURNARAKIS M， AMJAD R A， et al. A white paper on neural network quantization［DB/OL］. arXiv preprint： 2106.08295， 2021.
[24]	ZHAO X Y， HUANG P， SHU X B. Wavelet-attention CNN for image classification［J］. Multimedia Systems， 2022， 28（3）： 915-924.
[25]	FINDER S E， AMOYAL R， TREISTER E， et al. Wavelet convolutions for Large receptive fields［C］∥Computer Vision-ECCV 2024. Cham： Springer， 2025： 363-380.
[26]	王晓柱，钮赛赛，张凯，等. 基于小波变换与特征提取的红外弱小目标图像融合［J］. 西北工业大学学报， 2020， 38（4）： 723-732.
	WANG X Z， NIU S S， ZHANG K， et al. Image fusion of infrared weak-small target based on wavelet transform and feature extraction［J］. Journal of Northwestern Polytechnical University， 2020， 38（4）： 723-732 （in Chinese）.
[27]	REDMON J， DIVVALA S， GIRSHICK R， et al. You only look once： Unified， real-time object detection［C］∥2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2016： 779-788.
[28]	REDMON J， FARHADI A. Yolov3： An incremental improvement［DB/OL］. arXiv preprint： 1804.02767， 2018.
[29]	WANG C Y， BOCHKOVSKIY A， LIAO H M. YOLOv7： trainable bag-of-freebies sets new state-of-the-art for real-time object detectors［C］∥2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway： IEEE Press， 2023： 7464-7475.
[30]	FINDER S E， AMOYAL R， TREISTER E， et al. Wavelet convolutions for Large receptive fields［C］∥Computer Vision-ECCV 2024. Cham： Springer， 2025： 363-380.
[31]	PAN J H， HE C， HUANG W， et al. Wavelet tree transformer： Multihead attention with frequency-selective representation and interaction for remote sensing object detection［J］. IEEE Transactions on Geoscience and Remote Sensing， 2024， 62： 5637023.
[32]	GONG R H， LIU X L， JIANG S H， et al. Differentiable soft quantization： Bridging full-precision and low-bit neural networks［C］∥2019 IEEE/CVF International Conference on Computer Vision （ICCV）. Piscataway： IEEE Press， 2019： 4851-4860.
[33]	HUANG L， DONG Z W， CHEN S L， et al. HQOD： Harmonious quantization for object detection［C］∥2024 IEEE International Conference on Multimedia and Expo （ICME）. Piscataway： IEEE Press， 2024： 1-6.

模型	算法	Pre/%	Rec/%	mAP₅₀/%	mAP_50∶95/%	Params/M	GFLOPs
A	YOLOv8s	52.4	41.4	40.8	24.3	11.14	28.70
B	YOLOv8s+MV2Block	47.2	37.5	35.8	20.4	9.51	20.93
C	YOLOv8s+SV2Block	46.9	37.1	35.5	20.1	7.92	18.65
D	YOLOv8s+GhostConv	46.2	36.7	35.1	19.4	5.93	16.27
E	D+W-DWConv	48.4	38.8	37.6	22.0	5.93	16.30
F	D+W-PWConv	48.0	38.4	36.7	21.3	6.04	16.47
G	D+W-DSConv	50.8	38.8	38.4	22.4	6.04	16.49

N	输入分辨率：1 024×1 024				输入分辨率：640×640				Params/M	GFLOPs
N	Pre/%	Rec/%	mAP₅₀/%	mAP_50∶95/%	Pre/%	Rec/%	mAP₅₀/%	mAP_50∶95/%	Params/M	GFLOPs
	46.2	36.7	35.1	19.4	40.6	29.9	28.5	15.7	5.93	16.27
1	48.3	37.9	37.1	21.7	41.9	32.1	29.8	17.0	5.93	16.28
2	48.0	38.7	37.4	21.9	42.3	31.9	30.1	17.2	5.93	16.29
3	48.4	38.8	37.6	22.0	42.9	32.5	30.4	17.4	5.93	16.30
4	48.2	38.4	37.3	21.8	42.4	32.2	30.2	17.2	5.93	16.31
5	47.9	37.7	36.9	21.6	42.1	32.0	29.9	17.1	5.93	16.32
6	47.6	37.4	36.6	21.3	41.8	31.8	29.7	16.9	5.93	16.34

M	输入分辨率： 1 024×1 024				输入分辨率： 640×640				Params/M	GFLOPs
M	Pre/%	Rec/%	mAP₅₀/%	mAP_50∶95/%	Pre/%	Rec/%	mAP₅₀/%	mAP_50∶95/%	Params/M	GFLOPs
	46.2	36.7	35.1	19.4	40.2	31.2	28.5	15.4	5.93	16.27
1	47.0	37.6	36.2	20.8	43.1	31.7	29.8	16.7	6.04	16.45
2	48.0	38.4	36.7	21.3	43.6	32.0	30.3	17.2	6.04	16.49
3	47.5	38.1	36.4	21.1	43.2	31.4	30.1	17.1	6.04	16.49
4	47.9	37.8	36.6	21.2	41.7	31.9	30.0	16.9	6.04	16.49
5	47.6	37.9	36.3	21.0	41.4	31.7	29.7	16.5	6.04	16.49
6	47.3	37.4	36.1	20.6	41.0	31.3	29.2	16.1	6.04	16.50

原始模型	mAP₅₀/%	mAP_50∶95/%	Params/M	GFLOPs	轻量化模型	mAP₅₀/%	mAP_50∶95/%	Params/M	GFLOPs
YOLOv8n	36.1	21.1	3.01	8.20	WaveYOLOv8n	32.4	18.8	1.72	5.23
YOLOv8n-p2	38.1	22.4	2.93	12.38	WaveYOLOv8n-p2	35.6	20.7	1.61	8.86
YOLOv8n-p6	36.5	21.5	4.79	8.19	WaveYOLOv8n-p6	33.4	19.3	2.70	5.24
YOLOv8s	40.8	24.3	11.14	28.70	WaveYOLOv8s	38.4	22.4	6.04	16.49
YOLOv8s-p2	42.3	25.2	10.64	36.97	WaveYOLOv8s-p2	39.8	23.3	5.43	22.53
YOLOv8s-p6	41.0	24.4	17.88	28.58	WaveYOLOv8s-p6	39.5	23.0	9.58	16.48
YOLOv5s	39.7	23.5	9.13	24.06	WaveYOLOv5s	38.6	22.4	5.98	16.58
NanoDet	28.5	15.6	0.94	3.61	WaveNanoDet	29.7	16.3	0.95	3.97

模型	比特设置	mAP₅₀/%	mAP_50∶95/%	Params/M	Sizes/MB	GFLOPs
YOLOv8s	FP32	40.8	24.3	11.14	44.56	28.70
WaveYOLOv8s	FP32	38.4	22.4	6.04	24.16	16.49
	W6A6	36.8	21.7	6.05	4.53	3.09
	W4A4	35.9	20.6	6.05	3.02	2.06
	W4A3	34.3	19.4	6.05	3.02	1.80
YOLOv8s-p2	FP32	42.3	25.2	10.64	42.56	36.97
WaveYOLOv8s-p2	FP32	39.8	23.3	5.43	21.72	22.53
	W6A6	38.1	22.5	5.44	4.07	4.22
	W4A4	37.7	21.8	5.44	2.72	2.82
	W4A3	36.2	20.1	5.44	2.72	2.46
YOLOv8s-p6	FP32	41.0	24.4	17.88	71.52	28.58
WaveYOLOv8s-p6	FP32	39.5	23.0	9.58	38.32	16.48
	W6A6	37.6	21.8	9.59	7.19	3.09
	W4A4	36.4	21.0	9.59	4.79	2.06
	W4A3	34.3	19.5	9.59	4.79	1.80

Wavelet time-frequency localization-based model compression for UAV object detection

RichHTML

PDF (PC)

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 16

References 33

Related Articles 15

Recommended Articles

Metrics

Comments

算法	比特设置	mAP₅₀/%	mAP_50∶95/%
PACT	W4A4	33.3	19.1
DSQ		29.9	16.6
LSQ+		33.6	19.5
N2UQ		32.7	18.4
HDOQ		34.2	20.1
W-FDQ（本文）		35.9	20.6
PACT	W4A3	28.4	16.1
DSQ		24.8	13.6
LSQ+		26.2	14.7
N2UQ		27.3	15.8
HDOQ		29.5	16.7
W-FDQ（本文）		34.3	19.4

[1]	Qishuai DING, Bangjun LEI, Zhengping WU. A lightweight single object tracking algorithm for UAVs based on Siamese network [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(4): 330925-330925.
[2]	Yiquan WU, Kang TONG. Research advances on deep learning-based small object detection in UAV aerial images [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(3): 30848-030848.
[3]	Shuai ZHONG, Liping WANG. MCS-RETR: Improved RT-DETR object detection method for UAV aerial images [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(22): 331987-331987.
[4]	Jihong ZHU, Jiacheng HAN, Xiaojun GU, Yahui ZHANG, Jun WANG, Jie HOU, Weihong ZHANG. Advances and challenges in cross-domain vehicle structures and morphing configuration design technologies [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(18): 431686-431686.
[5]	Yi ZHENG, Xianghong CHENG, Xingbang TANG, Yi CAO. Oriented detection algorithm for insulator and their defects from aerial images based on improved ReDet [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(18): 331825-331825.
[6]	Lin CHEN, Qing ZHU, Han HU, Yulin DING, Pengxin GU. FLASH: Flexible and lightweight awareness of slope hazard [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(11): 531293-531293.
[7]	Fanteng MENG, Yong QIN, Jing CUI, Yunpeng WU, Zicheng ZHANG, Shaowei WEI. Unknown risk detection in external environment of railroad using UAV images [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(11): 531262-531262.
[8]	Shusheng CHEN, Muliang JIA, Jiahao LIN, Shiyi JIN, Zhenghong GAO, Yueqing WANG, Zhiqiang MA, Zheng LI, Chenlong DUAN, Jiawei LI. Empowering aircraft technology applications with generative models: Research progress and prospects [J]. Acta Aeronautica et Astronautica Sinica, 2025, 46(10): 631194-631194.
[9]	Xudong LUO, Yiquan WU, Jinlin CHEN. Research progress on deep learning methods for object detection and semantic segmentation in UAV aerial images [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(6): 28822-028822.
[10]	Rui SI, Yong CHEN. Application trends of additive manufacturing technology for civil aircraft [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(5): 529677-529677.
[11]	Weihong ZHANG, Changhong TANG. Lightweighting of aerospace and aeronautical equipment: Challenges and perspectives [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(5): 529965-529965.
[12]	Zi WANG, Jinghao WANG, Yang LI, Zhang LI, Qifeng YU. Non-cooperative target pose estimation from monocular images based on lightweight neural network [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(22): 330248-330248.
[13]	Yixuan YOU, Xinchun JI, Dongyan WEI, Yi LU, Hong YUAN. Criterion MP-S for multi-scale cascaded geomagnetic matching similarity measurement [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(21): 330149-330149.
[14]	Wenyu WANG, Feng LI, Feixiang REN, Xingyu WEI, Jian XIONG. Research progress on structural design methods and mechanical properties of lightweight high⁃strength composite lattice stiffened shell structure [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(17): 530001-530001.
[15]	Junyu LI, Qiankun LIU, Ying FU. Infrared small object detection based on attention mechanism [J]. Acta Aeronautica et Astronautica Sinica, 2024, 45(14): 628959-628959.