基于浅层卷积神经网络的实时视频通信隐私保护研究  

On Privacy Protection of Real-time Video Communication Based on Shallow Convolution Neural Network

作　　者：陈晨 张自军 姚保峰 CHEN Chen;ZHANG Zijun;YAO Baofeng(School of Computer engineering,Bengbu University,Bengbu 233030,China)

机构地区：[1]蚌埠学院计算机工程学院,安徽蚌埠233030

出　　处：《宿州学院学报》2020年第3期80-84,共5页Journal of Suzhou University

基　　金：蚌埠学院校级科研项目(2015ZR07);蚌埠学院工程中心研究项目(BBXYGC2014B04);国家级大学生创新创业计划项目(201511305008)。

摘　　要：基于tiny-YOLOv3算法,提出了一种浅层网络的目标检测卷积神经网络结构,用于实时视频通信中的隐私保护。此网络引入特征金字塔的思想,输出两个尺度的预测张量,改善小目标检测的结果。引入全卷积神经网络思想,在两个尺度特征图上进行上采样,在目标检测的同时并行进行语义分割,生成掩模原型。引入前景人员和背景人员的概念,在目标检测阶段只检测出背景人员。最后将语义分割和目标检测的结果进行组合,将检测到的背景人员进行遮挡从而实现实时视频通信中的隐私保护。实验表明,此网络在性能一般的机器上实现了准确率和帧数的较好平衡,可以应用于实时视频通信的隐私保护。Based on tiny-YOLOv3 algorithm,this paper presents a shallow target detection convolution neural network,which is used for privacy protection in real-time video communication.This network introduces the idea of feature pyramid,outputs two scales of predictive tensor,and improves the results of small target detection.The idea of full convolution neural network is also introduced to sample on two scales of feature maps.Semantic segmentation is carried out simultaneously with target detection to generate mask prototype.Introducing the concept of foreground person and background person,only background person are detected during the target detection phase.The results of semantic segmentation and target detection are combined finally to block the detected background person so as to realize the privacy protection in real-time video communication.Experiments show that the network realizes the good tradeoff of accuracy and FPS on the performance-based machine,and can be applied to the privacy protection of real-time video communication.

关 键 词：视频隐私保护 目标检测 图像分割 YOLO 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]