摘要
针对目前目标检测技术中小目标检测困难问题,提出了一种基于SSD (Single Shot multibox Detector)改进的小目标检测算法Bi-SSD (Bi-directional Single Shot multibox Detector).该算法为SSD的浅层特征设计了小目标特征提升模块,在网络的分类和回归部分结合多尺度特征融合方法和BiFPN (Bi-directional Feature Pyramid Network)结构,设计了6尺度BiFPN分类回归子网络.实验结果表明,在PASCAL VOC和MS COCO目标检测数据集上Bi-SSD相比原始的SSD算法有更好的检测性能.其中VOC2007+2012上Bi-SSD算法的mAP指标达到了78.47%相较SSD算法提升了1.34%,在COCO2017上Bi-SSD算法的m AP达到26.4%提升了接近2.4%.
Aiming at the difficulty of small target detection in current target detection technology,a small target detection algorithm named improved Bi-directional Single Shot multibox Detector(Bi-SSD)based on Single Shot multibox Detector(SSD)is proposed.This algorithm designed a small object feature improvement module for the shallow features of SSD.In the classification and regression parts of the network,a 6-scale Bi-directional Feature Pyramid Network(BiFPN)is designed as classification and regression sub-network according to multi-scale feature fusion method and BiFPN structure.Experimental results show that Bi-SSD has better detection performance than the original SSD on PASCAL VOC and MS COCO object detection datasets.On VOC2007+2012,Bi-SSD achieves 78.47%mAP,which is an increase of 1.34%compared to the original SSD algorithm.On COCO2017,Bi-SSD achieves 26.4%mAP,which was an increase of 2.4%compared to the original SSD algorithm.
作者
汪能
胡君红
刘瑞康
范良辰
WANG Neng;HU Jun-Hong;LIU Rui-Kang;FAN Liang-Chen(College of Physical Science and Technology,Central China Normal University,Wuhan 430079,China)
出处
《计算机系统应用》
2020年第11期139-144,共6页
Computer Systems & Applications
关键词
小目标检测
特征融合
多尺度检测
深度学习
卷积神经网络
small object detection
feature fusion
multi-scale detection
deep learning
Convolutional Neural Network(CNN)
