Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition

URL:http://ydwen.github.io/papers/WenECCV16.pdf
这篇论文主要的贡献就是提出了Center Loss的损失函数,利用Softmax Loss和Center Loss联合来监督训练,在扩大类间差异的同时缩写类内差异,提升模型的鲁棒性。

Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition
为了直观的说明softmax loss的影响,作者在对LeNet做了简单修改,把最后一个隐藏层输出维度改为2,然后将特征在二维平面可视化,下面两张图分别是MNIDST的train集和test集,可以发现类间差异比较明显,但是类内的差异也比较明显。
Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition
为了减小类内差异论文提出了Center Loss:
Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition大专栏  Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition-Deep-Face-Recognition-image004.png" alt=""/>
Cyi就是类的中心点特征,Cyi的计算方法就是yi类样本特征的均值,为了让center loss在神经网络训练过程中切实可行,Cyi的计算是对于每一个mini-batch而言,因此结合Softmax Loss,整个网络的损失函数就变成了, λ用来平衡这两个Loss:
Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition
用同样的网路结构只是将Softmax Loss替换成Center Loss作者在MNIST数据集上做了同样的实验,对于不同的λ值得到了如下可视化结果可以发现Center Loss还是比较明显的减小了类内差异同时类间差异也比较突出。
Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition
在公开数据集上的表现:
Center Loss - A Discriminative Feature Learning Approach for Deep Face Recognition

上一篇:[论文阅读] A Discriminative Feature Learning Approach for Deep Face Recognition (Center Loss)


下一篇:East Central North America Region 2015