计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (6): 1566-1578.DOI: 10.3778/j.issn.1673-9418.2403029

• 图形·图像 • 上一篇    下一篇

特征级自适应增强的无人机目标检测算法

颜豪男,吕伏,冯永安   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 出版日期:2024-06-01 发布日期:2024-05-31

Feature-Level Adaptive Enhancement for UAV Target Detection Algorithm

YAN Haonan, LYU Fu, FENG Yong'an   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2024-06-01 Published:2024-05-31

摘要: 在自然光照条件下,无人机航拍图像中包含着低照度图像,这部分低照度图像会影响目标检测的精度,并且基于航拍图像的检测任务往往具有较高的实时性需求。针对上述问题,提出了一种特征级自适应增强的无人机航拍图像目标检测算法。首先使用改进的拉普拉斯算子与IAT图像增强网络进行融合,以强化目标的边缘特征,提升目标检测能力;其次使用双分支结构对正常图像与增强图像进行并行学习,以自适应选择的方式对学习到的特征进行融合,从而构建BLENet特征级自适应增强网络,自动适应光照并增强图像信息;然后设计基于频域分离的可变形卷积FS-DC模块以及具有参数共享的FS-C2F模块,从而在增强高频信息捕捉能力的同时减少模型参数量和计算冗余;最后改进回归损失函数Wise-IoU,使模型进一步聚焦于中高质量锚框,从而降低边界回归损失,提升定位精度。在公开数据集visDrone2023上的实验结果表明,相较于基线模型,最终模型在保持99 FPS的前提下,mAP@0.50提升了2.3个百分点,适用于露天光照变化环境下的无人机航拍实时检测任务。

关键词: 无人机(UAV), 目标检测, 图像增强, YOLOv8, 低照度, 自适应, 频域分离

Abstract: Under natural lighting conditions, UAV (unmanned aerial vehicle) aerial images contain low-light images, which affect the accuracy of target detection, and detection tasks based on aerial images often have high real-time requirements. To address the above problems, a feature-level adaptive enhancement target detection algorithm for UAV aerial images is proposed. Firstly, an improved Laplace operator is fused with the IAT (illumination adaptive transformer) image enhancement network to enhance the edge features of the target and improve the target detection ability. Secondly, a two-branch structure is used to learn the normal image and the enhanced image in parallel, and the learnt features are fused in an adaptive selection manner to construct a BLENet (bad lighting enhancement net) feature-level adaptive enhancement network, which can adapt to the illumination and enhance the information of the image automatically. Then, the deformable convolutional FS-DC module based on frequency domain separation and the FS-C2F module with parameter sharing are designed so as to reduce the number of model parameters and computational redundancy while enhancing the ability of capturing high-frequency information. Finally, the regression loss function Wise-IoU is improved so that the model further focuses on the medium- and high-quality anchor frames, thus reducing the boundary regression loss and improving the localization accuracy. Experimental results on the publicly available dataset visDrone2023 show that compared with the baseline model, the final model improves the mAP@0.50 by 2.3 percentage points while keeping the FPS at 99 frames per second, which makes it suitable for UAV aerial photography real-time inspection and monitoring tasks under the environment of varying light in the open air.

Key words: unmanned aerial vehicle (UAV), object detection, image enhancement, YOLOv8, low illumination, adaptive, frequency domain separation