Abstract. Convolutional neural networks (CNN) have immensely promoted the development of face recognition (FR) technology. In order to achieve global accuracy, CNN models tend to be deeper or multiple local facial patch ensembles, leading to excessive amounts of calculation. We address these deep FR problems and propose a lightened deep learning framework under an open-set protocol to achieve a good classification effect and streamline the model itself. To this end, we improve the SphereFace that enables the deep network to learn angularly discriminative features more efficiently. First, global average pooling (GAP) is introduced to replace the original fully connected (FC) layer, which greatly reduces the storage of the model. Compared to the widely used FC layer, GAP can reduce the number of parameters and avoid overfitting. Then multilayer perceptron is added between convolution layers, which increases the ability to characterize features. These models are trained on the CASIA-WebFace dataset and evaluated on the Labeled Faces in the Wild and YTF datasets, which show the excellent performance of lightened SphereFace (L-SphereFace) in FR tasks. At the same time, computational cost is reduced in comparison with the released SphereFace model. The storage space of the model is also greatly compressed.

中文翻译：

---

用于人脸识别的 Lightened SphereFace

摘要。卷积神经网络（CNN）极大地促进了人脸识别（FR）技术的发展。为了实现全局精度，CNN 模型往往更深或多个局部面部补丁集成，导致计算量过大。我们解决了这些深度 FR 问题，并在开放集协议下提出了一个轻量级的深度学习框架，以实现良好的分类效果并简化模型本身。为此，我们改进了 SphereFace，使深度网络能够更有效地学习角度判别特征。首先，引入全局平均池化（GAP）来代替原来的全连接（FC）层，大大减少了模型的存储。与广泛使用的 FC 层相比，GAP 可以减少参数数量并避免过拟合。然后在卷积层之间加入多层感知器，增加了表征特征的能力。这些模型在 CASIA-WebFace 数据集上进行了训练，并在 Wild 和 YTF 数据集中的 Labeled Faces 上进行了评估，这显示了 Lightened SphereFace (L-SphereFace) 在 FR 任务中的出色表现。同时，与发布的 SphereFace 模型相比，降低了计算成本。模型的存储空间也大大压缩。与发布的 SphereFace 模型相比，计算成本降低。模型的存储空间也大大压缩。与发布的 SphereFace 模型相比，计算成本降低。模型的存储空间也大大压缩。