| dc.description.abstract | With the progression of technology, the demand for life automation has become more and more popular, and face recognition has become an indispensable part of various automation applications and has been widely used in many different fields. In recent years, the development of deep learning is booming, in order to pursue the accuracy, face recognition has also introduced a large number of deep learning techniques. The network models were modified deeper and deeper, the number of parameters were continuously increased, and the requirements for hardware specifications were then upgraded again and again. On the other hand, it is hard to obtain face photos due to the privacy issues.
To solve the problems, we propose a few-shot face recognition system based on a lightweight convolutional neural network, and mainly study how to improve the accuracy of face recognition in the case of lightweight models and few-shot learning. At first, we take MobileFaceNet as the backbone and remove the final 11 convolutional layer, thereby reducing the model size, and added ECA-Net modules to maintaining the accuracy. We call this new model ECA-MobileFaceNet-S. We can input face photos into the model to obtain the feature vectors, and then classify according to their cosine similarity. Then, to improve the few-shot learning problems, we refer to Prototypical Networks and replace the square Euclidean distance with large angular margin loss to improve the face recognition ability.
We reduce the parameters of the original MobileFaceNet by 13.39%. Compared to the Prototypical Networks, the accuracy on the LFW dataset is improved by 10.79% under the effect of ProtoNets training method and the large angular margin loss. | en_US |