部分卷积耦合双通道网络的低分辨率人脸识别

Partial Coupled Convolution Layers in Dual Channel Network for Low Resolution Face Recognition

在实际应用场景中所采集的人脸图像通常分辨率都很低,导致许多经典的人脸识别算法无法对低分辨率人脸进行准确识别.针对该问题,本文提出了一种部分卷积耦合的双通道网络,该网络中将高分辨率(High Resolution,HR)通道和低分辨率(Low Resolution,LR)通道中的卷积核进行部分耦合,使得LR通道能够从耦合的卷积核中学习到HR通道中的高分辨率参数,从而达到提高LR通道对LR样本的特征提取能力.为进一步提高样本分类的准确率,在双通道网络末端引入一个空间金字塔池化层(Spatial Pyramid Pooling,SPP),使用SPP层能够将HR样本与LR样本投影到一个共同的特征空间中.最后使用LFW人脸库对算法的有效性进行测试,实验结果表明,本文所提算法能够对LR人脸图像进行准确识别.

Face images collected in most realistic scenarios are usually low resolution,which makes many classic face recognition algorithms unable to accurately recognize these low resolution face images.To solve this problem,this paper proposes a partially coupled convolution layers in dual-channel network in which the convolution kernels in the high resolution(HR)channel and the low resolution(LR)channel are partially coupled.In this way,the LR channel can learn the high-resolution parameters of the HR channel from the coupled convolution kernel,to enhance the feature extraction ability of the LR channel.In order to further improve the accuracy of sample classification,the paper introduces a spatial pyramid pooling(SPP)layer at the end of dual-channel network and uses the SPP layer to project HR samples and LR samples into a common feature space.Finally,the LFW face database is used to test the effectiveness of the proposed algorithm.The experimental results show that the algorithm proposed in this paper can accurately recognize LR face images.