High Performance Face Identification System with Optimized Deep Learning Architectures

Abstract

Human identification is an essential element in authentication systems and surveillance devices. Various technologies are utilizing this method to perform higher-level processing and provides additional values in their solutions such as in smartphone, automatic gate, attendance recorder, social media, and smart CCTV system.

The work on this manuscript focus on the camera-based identification system in a real-world scenario where human faces are the main feature. This kind of system requires a fast processing time as well as acceptable identification quality. To achieve these two requirements, the research in this study proposes efficient deep learning models for detecting and identifying faces from a given video stream.

A complete face identification system consists of face detection and face identification system. For the face detection module, the proposed model uses an altered version of the single-shot detection model. This model is optimized to detect not only large faces but also the small ones. By utilizing the low-level and high-level features, the model can deliver satisfying performance. Furthermore, an attentionbased mechanism is incorporated to helps the training process of the model.

As the detection module predicts un-aligned faces, the identification result is compromised. Therefore, an alignment procedure is needed to rectify the predicted faces. Instead of using a separated model, this module is embedded directly in the face detection module. This kind of strategy allows the system works efficiently. Compared to the two-stages approach, where the detection and alignment process uses separated models, this strategy works more efficiently.

The face identification module extracts an embedding vector from a given face image. This module utilizes a deep learning model that extract distinctive features for each identity. It is achieved by train the model using the angular margin strategy to ensure the feature vector that corresponds to an identity form a distinct cluster in the embedding space. The model is then augmented using the auto-encoder principle to fortify the feature transformation result.

Despite the satisfying performance, the deep learning model often suffers in processing time. A model needs more layers to achieve better results, which leads to more processing time. Therefore, this study utilizes a channel pruning strategy to alleviate this problem. As a result, significant improvement in the processing speed is achieved, allowing the whole pipeline of the face identification model to be executed efficiently for real-world applications.