航空学报 > 2025, Vol. 46 Issue (23): 631952-631952   doi: 10.7527/S1000-6893.2025.31952

干扰环境下无人机多源感知专栏

小波时频局部化无人机目标检测模型压缩研究

黄维, 潘家皓, 何楚()   

  1. 武汉大学 电子信息学院,武汉 430072
  • 收稿日期:2025-03-10 修回日期:2025-03-28 接受日期:2025-04-25 出版日期:2025-05-23 发布日期:2025-05-08
  • 通讯作者: 何楚 E-mail:chuhe@whu.edu.cn
  • 基金资助:
    国家重点研发计划(2016YFC0803000);国家自然科学基金(41371342)

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

Wei HUANG, Jiahao PAN, Chu HE()   

  1. School of Electronic Information,Wuhan University,Wuhan 430072,China
  • Received:2025-03-10 Revised:2025-03-28 Accepted:2025-04-25 Online:2025-05-23 Published:2025-05-08
  • Contact: Chu HE E-mail:chuhe@whu.edu.cn
  • Supported by:
    National Key Research and Development Program of China(2016YFC0803000);National Natural Science Foundation of China(41371342)

摘要:

基于深度学习的遥感目标检测模型依托地面服务器强大的算力与存储资源,已具备高效处理海量遥感数据的能力。然而,在无人机等移动边缘设备部署场景中,受限于计算资源、存储容量与能耗约束,大规模模型难以实现有效部署。模型压缩,如轻量化设计、模型量化等,已成为推动遥感目标检测算法落地的重要技术。以单阶段的目标检测深度网络为基准,提出一种基于小波时频局部化的模型压缩框架:通过深度可分离卷积与离散小波变换的空频局部化特性相融合,构建具有扩展感受野的小波深度可分离卷积(W-DSConv)模块,实现了模型的轻量化重构;基于小波频域分解特性,提出小波分频量化(W-FDQ)的量化感知训练方法,实现不同频段特征的独立量化,进一步完成对轻量化模型的压缩。实验中选取了YOLO系列的网络模型进行验证,在VisDrone2021无人机遥感数据集上的实验结果表明:所提W-DSConv模块在参数量减少45.5%、计算量减少37.4%的情况下,检测精度波动幅度控制在2.2%以内;采用W-FDQ方法进行6 bit与4 bit量化时,量化模型分别保持了浮点模型的95.8%与92.9%的检测性能。为移动端遥感目标检测模型的轻量化部署提供了新的技术思路。

关键词: 目标检测, 无人机, 模型压缩, 轻量化, 模型量化, 小波变换

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

中图分类号: