| 摘要: |
| 无人机目标检测与识别任务中，目标随着飞行高度的改变尺寸发生显著变化。常规目标检测模型中，获取的小目标细节信息有限，检测精度较低；而适用于小目标的实时检测模型往往容易丢失大目标的背景信息，降低大目标的检测精度。针对以上多尺度目标检测识别任务难点，提出一种基于改进特征金字塔网络(Feature Pyramid Network,FPN)结构的实时多尺度目标检测识别模型。该模型通过增加特征金字塔层级覆盖更广的目标尺度，获取更为丰富的目标信息；同时，利用跨连接增加不同尺度特征融合的多样性，降低特征传导距离，保留更加完整的尺度特征来提高模型检测识别多尺度目标的性能。通过实验发现,相比于原始网络结构和相同特征层级的四层特征金字塔结构，加入改进特征金字塔结构的多尺度目标检测模型识别性能得到了提升。 |
| 关键词: 无人机(UAV) 目标检测与识别 多尺度目标检测 特征金字塔网络（FPN） |
| DOI： |
|
| 基金项目: |
|
| A real-time multi-scale object detection and identification algorithm |
| ZHU Peipei,WU Yuan,LAI Zuomei |
| (Southwest China Institute of Electronic Technology,Chengdu 610036,China) |
| Abstract: |
| In unmanned aerial vehicle(UAV) object detection and identification task,targets’ sizes vary enormously alone with the height of the UAV.Some normal object detection and identification algorithms can only extract limited small object features,which results in low detection performance.Researchers also develop some deep neural networks for small objects.But these models always have small receptive field and lose some large targets’ features.In this paper,a real-time object detection and identification network based on improved Feature Pyramid Network(FPN) module is proposed to promote the performance of multi-scale objects detection.This improved FPN module increases the feature hierarchies to capture more object information.Moreover,the cross connection is used in the modified FPN to reduce feature conduction path for reserving useful features.The experiment demonstrates that the performance of object detection and identification in two public datasets composed of aerial photos is increased compared with initial networks and the four-layer FPN. |
| Key words: unmanned aerial vehicle(UAV) object detection and identification multi-scale object detection feature pyramid network(FPN) |
