Face Recognition based on Deep Learning

Face Recognition (FR) is currently applied in two of Eliteun systems: Lightweight IoT Access Control System (LIACS) and Multi-factor Vehicle Surveillance System (MVSS). The storage of a single FR camera can be extended to 50,000 photos. With great capacity and fast processing, the technology will be applied in more IoT scenarios to fulfill the requirements for lightweight, easy to deploy and instant response.

Face recognition has a longer history with more developed technologies and more realization methods. In the earlier period, manual modeling was popular, but engineers had to handcraft all features for further coding and classification. The development of cloud computing and big data reveals manual model’s inadequacy for fast process of mass data. Thus deep leaning is applied to optimize face recognition. Deep learning is part of a broader family of machine learning methods, which identifies data distribution characteristics and learns via artificial neural networks. Features in an image are extracted and filtered through multiple layers in CNN (Pic.1) This automatic model remarkably simplifies the process of complex classification, so the efficiency rises.

Basic CNN structure contains input, convolutional, pooling, fully connected layer and output layer (classifier). However, the application of CNN can differ greatly in modeling, training and optimizing in different scenarios. Eliteun team adopted OpenCV for preprocess of face images including face detection, grayscale and image resize. (Pic.2) Preprocessed images are sent to CNN for feature extraction, classification, and matching with existing data. Therefore, preprocess and modeling training both impact recognition accuracy and speed. Adopting neural networks in deep learning significantly simplifies the process, reduces the load on system and increases the speed.