Face re-identification in thermal infrared spectrum based on ThermalFaceNet neural network

Face recognition is widely explored research area which achieved high level of maturity. It attracts scientists to explore new possibilities and various spectral domains. Thermal infrared spectrum seems to be a promising modality which may complement visible domain systems. Face recognition process consists of several stages such as image acquisition, face detection, feature extraction and matching. Extraction of features is the one of most important stage that can be performed using various approaches, such as appearance-based methods, local descriptor methods or convolutional neural networks. Lately, convolutional neural networks (CNN) become very popular. CNN is a structure based on various filtering layers to reduce size, extract features and finally to classify the input data. This technique allows to efficiently process large datasets. Extracted features can be used to perform multiclass classification or identity verification. Verification which refers to comparing two samples is often performed using distance metrics. This paper presents thermal face verification method based on Siamese convolutional neural network. We introduce ThermalFaceNet architecture and compare performance with existing state-of-the-art CNN architectures.