面向无依托供电场景的小目标检测轻量级模型
Lightweight small object detection model for no reliable power supply scenes
冀金金 1荆有波2
作者信息
- 1. 郑州大学 河南先进技术研究院,河南 郑州 450002
- 2. 中国科学院微电子研究所通信与信息工程研发中心,北京 100029
- 折叠
摘要
针对现有目标检测模型在无依托供电场景存在检测效果不稳定、小目标大量漏检的问题,基于YOLOv4-tiny提出一种改进模型AMS-YOLOv4-tiny.通过在主干网之后引入更平滑的Mish函数、设计一种浅层特征加固的特征融合网络SCFPN、反复嵌入通道注意力机制3种策略,大幅提升预测特征层对目标的表达能力.实验结果表明,算法在PASCAL VOC07+12 数据集上的 mAP(mean of average precision)达到 87.19％,相比 YOLOv4-tiny 提高 4.45％,且部署在嵌入式设备上进行可行性验证,满足多种复杂场景下人车检测任务的精度与实时性要求.
Abstract
Aiming at the problems that the existing object detection models have unstable detection effects and a large number of small objects is misdetected in the scenario of no reliable power supply,the AMS-YOLOv4-tiny was proposed based on the YOLOv4-tiny.A smoother Mish function was introduced after the backbone network,a shallow feature consolidation of feature fusion network(SCFPN)was designed,and the channel attention mechanism was repeatedly embedded,the expression ability of the prediction feature layer on the object was then greatly improved.Results of experiments show that the mean average preci-sion(mAP)obtained by testing the AMS-YOLOv4-tiny algorithm on the PASCAL VOC07+12 data set is 87.19％,which is 4.45％higher than that of the YOLOv4-tiny.The AMS-YOLOv4-tiny algorithm was deployed on embedded devices for feasibility verification.Results show that it meets the accuracy and real-time requirements of human and car detection tasks in a variety of complex scenes.
关键词
小目标检测/特征融合网络/浅层特征加固/通道注意力机制/无依托供电/目标检测环境多变/低功耗Key words
small object detection/feature fusion network/consolidation of shallow features/channel attention mechanism/no reliable power supply/variable object detection environments/low power consumption引用本文复制引用
基金项目
国家重点研发计划基金项目(2018YFC0823900)
出版年
2024
NETL
NSTL
万方数据