TIRFaceNet: Thermal IR Facial Recognition
Samah A. F. Manssor, Shaoyuan Sun · 2019
Facial recognition using deep neural networks has attracted increasing attention in recent years. Although great progress has been made in facial recognition systems, there are many problems which hinder the identification process, especially the recognition under low light conditions or at night (full darkness). The goal of this study is to produce a model that identifies persons at night especially unauthorized persons entering a building and to assist the security workgroup in making appropriate decisions. To achieve this goal, we propose a deep convolutional neural network (CNN) based model called TIRFaceNet. Our method realizes pre-processing of thermal and visible images, offers unique features of the face (such as eyes, nose, chin, etc.) which are picked separately and train faces on the deep network. This model analyzes the deep features of a person's face and compares these features with other person's features stored in the dataset to recognize a person's face. In this paper, the proposed TIRFaceNet model achieves a better identification rate compared to other methods. The accuracy of our method is 98.70% ± 0.03 on the DHU database. In the DHUFO database, it is 98.50% ± 0.05. In addition, our method can identify a visible-to-thermal face with less training time (10 h <; 20 h) and with less processing time (0.02 s <; 0.025 s).