A Single-Stage Face Detection and Face Recognition Deep Neural Network

Based on Feature Pyramid and Triplet Loss

Abstract:

A practical deep learning face recognition system can be divided into several tasks. These tasks can be time-consuming if executed each task with the original image as the input data. And the feature extractors using by different task may duplicate its function. In this paper, we propose a multi-task training method with the optimization of deep neural network to train a single-stage deep neural network for face detection and recognition. We use the feature pyramid and triplet loss combined with anchor boxes to localize the face location, and the regression layer of the neural network to align the face position. This structure combines the multiple faces detection task and recognition task in a single network without a significant precision lose where it was usually implemented in multi-stage network. Alternatively, the feature extraction task and feature matching task can be combined to extract face features by using convolution and match the person identification by a full connection layer with softmax function. Also the proposed loss function is used to establish the feature extractor and finally match the features through a simple mathematical function. On an Nvidia 2080Ti GPU accelerator, this system can achieve 212 FPS for a 640x640 resolution input.

Model Architecture:

A comparison between typical two-stage model and proposed single-stage model

The proposed model architecture

Made by ¬ö¬f§Ê