面向算力受限边缘环境的双分支多尺度感知人脸检测网络  

Multi-scale aware dual path network for face detection in resource-constrained edge computing environment

作　　者：戚琦[1] 马迎新 王敬宇[1] 孙海峰[1] 廖建新[1] QI Qi;MA Yingxin;WANG Jingyu;SUN Haifeng;LIAO Jianxin(State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学网络与交换国家重点实验室,北京100876

出　　处：《通信学报》2020年第8期165-174,共10页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2018YFB1800502);国家自然科学基金资助项目(No.61671079,No.61771068);北京市自然科学基金资助项目(No.4182041)。

摘　　要：针对边缘算力受限,难以部署复杂结构的人脸检测深度神经网络的问题,为减少资源消耗,并保证人脸在多尺度变化、遮挡、模糊、光照等复杂场景下的检测精度,提出了多尺度感知的轻量化人脸检测算法。采用改进的人脸残差神经网络作为特征提取网络,并提出双分支浅层特征提取模块,并行分支理解图像多尺度信息,进而由深浅特征融合模块将底层图像信息与高层语义特征融合,配合多尺度感知的训练策略监督多分支学习差异化特征。实验结果表明,所提算法可有效提取多样化的特征,在保持模型高效性和低推理时延的同时,有效提升了算法的精度和稳健性。Aiming at the problem that face detectors with complex deep neural structures are difficult to deploy in the resource-constrained edge computing environment,to reduce the resource consumption while maintain the accuracy in complex scenes such as multi-scale face changes,occlusion,blur,and illumination,SDPN(multi-scale aware dual path network)for face detection was proposed.The Face-ResNet(face residual neural network)was improved,and a dual path shallow feature extractor was used to understand the multi-scale information of the image through parallel branches.Then the deep and shallow feature fusion module,a combination of the underlying image information and the high-level semantic feature,was used in conjunction with the multi-scale awareness training strategy to supervise the multi-branch learning discriminating features.The experimental results show that SDPN can extract more diversified features,which effectively improve the accuracy and robustness of face detection while maintaining the efficiency of the model and low inference delay.

关 键 词：人脸检测 多尺度感知 特征融合 人脸特征分析 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]