摘要
针对实时目标检测SSD(Sing shot multibox detector)算法对小目标检测能力弱的问题,提出一种提高特征图分辨率的超分辨率重建SRCNN(Super Resolution Convolutional Neural Networks)设计策略。改进算法是在SSD基础网络VGG_16网络的conv4_3卷积层上进行的,把conv4_3卷积层产生的特征图通过SRCNN网络进行超分辨重建以提高conv4_3卷积层的特征图分辨率。然后再利用超分辨重建后的特征图和原特征图一起为小目标检测提供所需要的特征。实验表明上述设计方法相比于原经典SSD算法具有更高的检测精度和检测能力,以及在小目标检测上的效果更加明显。
For the question which is the real-time target detection algorithm of Sing shot multi-box detector has little ability to detect small targets.A super resolution reconstruction strategy of Super Resolution Convolutional Neural Networks for improving the resolution of feature maps is proposed.The improved algorithm is carried out on the conv4_3 volume layer of the SSD basic network VGG_16 network.This paper uses the super resolution reconstruction of the feature map generated by the conv4_3 volume layer is carried out through the SRCNN network to improve the characteristic map resolution of the conv4_3 volume layer.Then,it uses the super resolution reconstruction feature map and the original feature map together to provide the required features for small target detection.Experiments show that the design method has higher detection accuracy and detection ability compared with the original SSD algorithm,and the effect of small target detection is more obvious.
作者
王烈
殷金伟
WANG Lie;YIN Jin-wei(Guangxi University,Guangxi Nanning 530004,China)
出处
《计算机仿真》
北大核心
2020年第3期430-434,共5页
Computer Simulation
基金
广西自然科学基金(2013GXNSFAA0019339)。
关键词
超分辨率重建
特征图
卷积层
目标检测
Super-resolution reconstruction
Feature map
Convolution layer
Object detection
作者简介
王烈(1969-),男(汉族),广西南宁市人,教授,硕士研究生导师,主要研究领域为FPGA和深度学习图像处理;殷金伟(1992-),男(汉族),河南周口人,硕士研究生,主要研究领域为FPGA和深度学习图像处理。
