AutoFace: How to Obtain Mobile Neural Network-Based Facial Feature Extractor in Less Than 10 Minutes?

Various mobile and edge devices have significantly different processing capabilities, making it challenging to develop a single universal architecture of a neural network to extract facial embeddings. In this paper, we study the automated machine learning techniques to design a neural network with the best performance on a concrete device. The novel procedure is proposed to choose the better subnetwork of the Supernet based on a genetic algorithm with a surrogate binary classifier to compare the expected accuracy of two subnetworks. The latter uses only encoding of a candidate subnetwork and does not require directly estimating its accuracy on a validation set. As a result, the most computationally efficient and accurate model in TensorFlow Lite format is obtained in less than 10 minutes for a specific device and latency constraint. An Android demo application has been developed to demonstrate the potential of designed neural networks. It is experimentally shown that the proposed approach is universal: it can extract deep embeddings for tasks such as face verification and facial expression recognition and for various types of devices, including smartphones and Raspberry Pi single-board mini-computers. Our models process one facial image in real-time and achieve much higher accuracy when compared to the best-known lightweight networks.