An IoT-based face recognition solution using a residual network model for deep metric learning

Abstract

Biometric identification has been widely used in recent years, mainly because they represent more secure authentication systems than conventional ones. In this context, facial recognition is highlighted since it allows detecting and recognizing a person in real-time for their facial characteristics. This technology is particularly important and used in many applications such as smart surveillance. The evolution in surveillance technologies, thanks to Internet of Things (IoT), allows greater automation of this process since many monitoring functions performed by people can be replaced by realtime recognition techniques, turning the system even smarter, giving more information to the user, or increasing security in monitoring environments. It is noted that society is at a point where different types of technologies are converging and adding up. It is known that computer vision techniques are being incorporated into surveillance systems and deep learning models have proven innovative in solving various visual recognition problems. In this sense, this dissertation proposes a surveillance system, which uses these techniques to identify the individuals present in the vision field of a camera through a combination including Histogram of Oriented Gradient (HOG), Support Vector Machine (SVM), and a deep learning model, called ResNet (Residual Network). The set of detection and recognition techniques was deployed in a hardware with limited processing power, quite common in IoT devices. The idea is to demonstrate that even under these conditions, the proposed architecture still manages to work with high precision and in real-time. To achieve the proposed objective, experiments were carried out in different scenarios to verify the accuracy and robustness of the techniques adopted under different conditions. Two techniques were used in the detection scenario, but only one was carried out in the experiments since it consumes 20 times less processing time when compared to the second. The accuracy of the ResNet model used reached about 99.38% in the Labeled Faces in the Wild (LFW) Benchmark while it manages to deliver a rate of 1-3 fps (frames per second), showing excellent results, especially considering an embedded system. The performance evaluation of the system against different types of noise showed high invariability with darkening of the images and high precision and robustness against blur type interference.