Real-time face verification on mobile devices using margin distillation

Face verification is an attractive yet challenging research area in computer vision. To make an improvement in the existing models for face verification, we proposed a CNN-based model for face verification on Mobile devices. The Multi-Task Convolutional Neural Network (MTCNN), as a pretrained model, is used for face detection. Some modifications are applied to the MobileFaceNet model and trained using the Margin Distillation cost function. To boost the model performance, the Dense Block and Depthwise separable convolutions are used in the model. Results on seven datasets confirm that the proposed model obtained the relative accuracy improvements of 0.017%, 1.384%, 0.483%, 0.124%, 2.185%, 0.684%, and 1.34%, compared to the baseline model, on the LFW, CPLFW, CPLFW, CFP FF, CFP FP, AGEDB_30, and VGG2_FP datasets, respectively. Furthermore, we collected a dataset, including a total of 4800 samples, with 80 sample images of 60 celebrities. Images are downloaded from Google Image Search. The proposed model obtained a verification accuracy of 99.760 on the collected dataset.