摘要
针对单目视觉目标检测,提出了一种基于single-stage深度学习的H_SFPN算法。该算法与现有的YOLOv3和CenterNet算法相比,在保证实时性能的条件下,可有效提高小目标检测的准确度。首先设计了一种新的网络架构(backbone),这种架构通过改进的沙漏(Hourglass)网络模型来提取特征图,以便充分利用底层特征的高分辨率以及高层特征的高语义信息。然后在特征图融合阶段提出了基于SFPN的特征图加权融合方法。最后,H_SFPN算法对目标位置和大小的损失函数进行了改进,可有效降低训练误差,并加快收敛速度。由MSCOCO数据集上的实验结果可知,所提H_SFPN算法明显优于Faster-RCNN,YOLOv3以及EfficientDet等现有的主流深度学习目标检测算法,其中对小目标的检测指标AP s最高,达到了32.7。
This paper proposes a single-stage deep learning based H_SFPN algorithm for monocular visual object detection.Compared with the existing YOLOv3 and CenterNet algorithms,the proposed algorithm can effectively improve the accuracy of small object detection without sacrificing the real-time performance.This paper designs a new network architecture(backbone),which uses an improved Hourglass network model to extract feature maps in order to make full use of the high resolution of the underlying features and the high semantic information of the high-level features.Then in the feature map fusion stage,a method SFPN based on the weighted fusion of feature maps is proposed.Finally,the proposed H_SFPN algorithm improves the loss function of the object position and size,which can effectively reduce the training error and accelerate the convergence speed.According to the experimental results on the MSCOCO data set,the proposed H_SFPN algorithm is significantly better than the existing mainstream deep learning object detection algorithms such as Faster-RCNN,YOLOv3 and EfficientDet.Among them,the small object detection index AP s of this algorithm is the highest,reaching 32.7.
作者
石先让
宋廷伦
唐得志
戴振泳
SHI Xian-rang;SONG Ting-lun;TANG De-zhi;DAI Zhen-yong(College of Energy and Power Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210001,China;Chery Advanced Engineering&Technology Center,Wuhu,Anhui 241006,China)
出处
《计算机科学》
CSCD
北大核心
2021年第4期130-137,共8页
Computer Science
基金
安徽省发改委重大研发项目。
关键词
深度卷积神经网络
目标检测
加权融合
网络架构
损失函数
Deep convolutional neural network
Object detection
Weighted fusion
Backbone
Loss function
作者简介
SHI Xian-rang,born in 1996,postgra-duate.His main research interests include autonomous driving,object detection and pattern recognition.nuaasxr@163.com;通信作者:SONG Ting-lun,born in 1965,Ph.D,professor,Ph.D supervisor.His main research interests include simulation driven vehicle architecture design and development,autonomous driving vehicles,and data driven energy management strategies for new energy vehicles.songtinglun@nuaa.edu.cn。
