摘要
针对采用大型卷积神经网络提取高维特征进行人脸识别时占用内存空间较大以及消耗大量计算资源的问题,提出一种结合全局与局部池化的深度哈希全卷积神经网络.第一,提出一种基于全局平均池化层的全卷积网络,用以减少网络参数以及压缩模型尺寸;第二,提出一种学习不同特征的融合损失方法,将哈希量化误差损失与分类损失进行加权融合,用以学习具有多分类性质的近似哈希编码.实验表明,该方法能够在Visual Geometry Group(VGG)框架下将识别效率提高68%,且准确率略有提升;融合损失方法扩展到Face Residual Network(Face-ResNet)框架时,在保持准确率的情况下将识别效率提高了23.7%。结果表明,该方法可在保证准确率的前提下有效地从特征提取和特征降维两方面提高识别效率,同时该方法还可扩展用于其他网络.
To reduce the memory occupancy rate and computational resources in face recognition with high-dimension features extracted from large convolutional neural networks,an efficient Foully Convolutional Network(FCN)of the deep)hash combined with global and local pooling.First,an FCN based on Global Average Pooling(GAP)is proposed to reduce network parameters and compress the model size.Second,a fusion method for learning approximate hash coding with multiple classification properties is used with Quantization Loss and Softmax Loss.Experimental results show that thcmcthod proposed can improve the efficiency up to 68%and that the Rank-1 accuracy is increased slightly with the Visual Geometry Group(VGG)framework.The fusion loss method can improve the efficiency up to 23.7%and the Rank-1 accuracy is maintained with the F'acc Residual Network(F^acc-RcsNct)framework.The results indicate that the proposed method can improve the efficiency both from feature extraction and reduction.tt also can be applied to other frameworks.
作者
曾燕
陈岳林
蔡晓东
ZENG Yan;CHEN Yuelin;CAI Xiaodong(School of Mechanical and Electrical Engineering,Guilin Univ.of Electronic Technology,Guilin 54 1004,China;School of Information and Comniunication,Guilin Univ.of Electronic Technology,Guilin 54 1004,China)
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2018年第5期163-169,共7页
Journal of Xidian University
基金
2016年广西物联网技术及产业化推进协同创新中心资助项目(WLW200601)
2016年广西科技计划资助项目(广西重点研发计划)(桂科AB16380264)
2016年"认知无线电与信息处理"省部共建教育部重点实验室基金资助项目(CRKL160102)
关键词
全局平均池化层
近似哈希编码
融合损失
全卷积网络
global average pooling
approximate hash coding
fusion loss
full convolution network
