基于联合损失函数的小规模数据人脸识别

小规模数据人脸识别的难点在于数据量少而变化多，直接用深度神经网络进行训练易出现过拟合现象.针对此问题，本文提出了基于联合损失函数的小规模数据人脸识别算法，即利用联合损失函数，在基于Softmax损失函数的大规模公开人脸数据集上得到的预训练模型上重新训练.该方法既能充分使用模型参数，也能够提高模型的特征表征能力.除此之外，本文中还使用了传统特征后处理方法进行对比评估，证明了该方法在小规模人脸数据集上的有效性.实验表明，本文方法能大幅度提高模型在学校新生人脸数据集的检索精度. 

Face Recognition of Small-Scale Dataset Based on Joint Loss Functions

In order to solve the problems induced in face recognition with small-scale datasets that small data size along with large changes and over-fit trend during directly training with deep neural networks, a face recognition method was proposed based on small-scale datasets with joint loss functions. This method was arranged to finetune a pre-trained model trained with large-scale public facial datasets based on Softmax loss function to make full use of all parameters in the model and improve feature representation capability of the model. Compared with conventional feature postprocessing methods, the effectiveness of this method was verified and evaluated. Experiment results show that this method can largely improve the performance of face retrieval on school freshmen face dataset.