A Light Convolutional Neural Network Architecture for Cross-Spectral Face Recognition

The rise in the utilization of multispectral imaging in surveillance, biometrics, and forensics has led to Cross-Spectral Face Recognition becoming a pertinent problem to solve. For instance, surveillance cameras switch from utilizing the Visible (VIS) spectrum during the day to the Near Infrared (NIR) spectrum at night, in order to capture footage in the dark. The process of matching face images captured on different sensing modalities is called Cross-Spectral Face Recognition (CSFR) also known as Heterogeneous Face Recognition (HFR). The variation in spectral characteristics between the NIR and VIS face images makes face recognition across these two information domains a challenging problem to solve. This adds to the complexity of building an efficient and accurate model that performs cross-spectral face recognition across the NIR and VIS domains. This paper proposes a light CNN model that performs cross-spectral face recognition across the NIR and VIS spectra. The proposed model has obtained higher accuracy scores when compared to denser Convolutional Neural Networks on CSFR across both spectra on various benchmark datasets such as CASIA NIR-VIS 2.0 (97.12%), Oulu-CASIA (99.85%), and HIT Lab2 (96.88%). Furthermore, this model attained the aforementioned accuracy scores with significantly fewer trainable parameters (0.102 million parameters).