基于特征重整及优化训练的遮挡人脸识别算法

Research on Blocking Face Recognition Algorithm based on Feature Restructuring and Optimization Training

以ResNet50网络为基础网络,提出了使用信息熵的权重特征对池化层的信息进行优化,在不增加额外参数的情况下使得网络特征提取更加充分;同时提出了改进型PSO(粒子群)算法对卷积网络误差反向传播阶段进行优化,通过比较GSA(引力搜索)优化前后最优粒子的适应度函数值,以此确保每次迭代中获得全局最优粒子.实验结果表明本文提出的方法有助于人脸的识别,尤其是在遮挡人脸的识别中,同时优化后的网络的训练收敛速度和识别速度也有了明显提升.

ResNet50 network based network is proposed, using information entropy weights to optimize the information of the pooling layer and make the network feature extraction more fully without additional parameters. An improved PSO is proposed, a particle swarm algorithm to optimize a convolution phase error back propagation network by comparing GSA(gravitational search) before and after optimization. The optimal particle fitness function value can ensure that each iteration obtains the global optimal particle. The experiment results show that the proposed method is useful in face recognition, especially in occlusion face recognition, and the optimized network training convergence speed and recognition rate are improved significantly.