航空学报 > 2024, Vol. 45 Issue (14): 630119-630119   doi: 10.7527/S1000-6893.2024.30119

无人机遥感图像实时小目标检测方法

刘延芳1,2(), 佘佳宇1, 袁秋帆3, 周芮1, 齐乃明1,2   

  1. 1.哈尔滨工业大学 航天学院,哈尔滨 150001
    2.哈工大苏州研究院,苏州 215104
    3.上海宇航系统工程研究所,上海 201109
  • 收稿日期:2024-01-08 修回日期:2024-01-18 接受日期:2024-03-26 出版日期:2024-07-25 发布日期:2024-04-10
  • 通讯作者: 刘延芳 E-mail:lyf04025121@126.com
  • 基金资助:
    国家自然科学基金(52272390);黑龙江省自然科学基金(YQ2022A009)

Real⁃time small target detection networks for UAV remote sensing

Yanfang LIU1,2(), Jiayu SHE1, Qiufan YUAN3, Rui ZHOU1, Naiming QI1,2   

  1. 1.School of Astronautics,Harbin Institute of Technology,Harbin 150001,China
    2.Suzhou Research Institute of HIT,Suzhou 215104,China
    3.Shanghai Aerospace System Engineering Institute,Shanghai 201109,China
  • Received:2024-01-08 Revised:2024-01-18 Accepted:2024-03-26 Online:2024-07-25 Published:2024-04-10
  • Contact: Yanfang LIU E-mail:lyf04025121@126.com
  • Supported by:
    National Natural Science Foundation of China(52272390);Natural Science Foundation of Heilongjiang Province(YQ2022A009)

摘要:

得益于深度学习方法的发展,近年来目标检测方法的性能有了很大的提升。然而,从无人机(UAV)遥感图像中检测目标仍然存在很大的挑战,原因包括:UAV遥感图像中目标分辨率小、背景复杂,现有算法难以满足实时性要求。面对这些挑战,提出了一种基于多尺度多深度特征提取(MMFE)网络的实时小目标检测(RTSTD)方法,能够高效的从UAV遥感图像中检测小目标。RTSTD将一幅输入图像剪裁成多个小尺寸的图像,并将一部分小尺寸图像输入到轻量化的MMFE网络中。因此,RTSTD具有处理任意分辨率的遥感图像而不丢失图像细节特征的能力。对于MMFE网络,提出了一种更有效的输出:重叠向量能够表示目标在输入图像中的位置和置信度。为了增强MMFE网络区分目标和复杂背景的能力,重新定义了正样本和负样本。为测试RTSTD的性能,从开源数据集UAV123、DTB70和AU-AIR中筛选重构了7个数据集,共8369张UAV遥感图像,涉及地面和海面场景下的小目标检测。结果证明,与现有的检测方法相比,RTSTD方法在准确性和速度方面都取得了改善,平均F1-Score>0.90,GPU运行每秒>66帧,CPU运行每秒>35帧。

关键词: 遥感图像, 小目标检测, 实时检测, 卷积神经网络, 特征融合

Abstract:

Benefiting from deep learning methods, the performance of object detection methods has greatly improved in recent years. However, significant challenges still exist in detecting targets from UAV remote sensing images. For example, the targets in UAV remote sensing images have small resolution and complex background, and the existing algorithms are difficult to meet the requirement for real-timeliness. To overcome these challenges, this paper proposes a Real-Time Small Target Detection (RTSTD) method based on a Multi-scalar & Multi-depth Feature Extraction (MMFE) network, which can efficiently detect small targets from UAV remote sensing images. The proposed RTSTD crops an input image into multiple small-size images, and feeds a portion of these small-size images into the lightweight MMFE network. Therefore, RTSTD has the capability to handle remote sensing images of arbitrary resolutions without losing image features. A more effective output is proposed for the MMFE network: an overlap vector that represents the position and confidence of the target in the input image. To enhance the MMFE network’s ability to distinguish targets from complex backgrounds, the positive and negative samples are redefined. To test the performance of RTSTD, seven datasets are selected and reconstructed from UAV123, DTB70 and AU-AIR, comprising a total of 8,369 UAV remote sensing images involving small target detection in the ground and sea scenarios. The experimental results demonstrate that compared to existing detection methods, the RTSTD method achieves improvements in both accuracy and speed. It achieves an F-Score of 0.90 or above, with a running speed of over 66 frames per second (FPS) using GPU acceleration and over 35 FPS using only CPU.

Key words: remote sensing image, small target detection, real-time detection, convolutional neural network, feature fusion

中图分类号: