Performance Evaluation of Deep Feature in Cross Spectral Face Recognition

Cross-spectral face recognition (CSFR), particularly in the Near Infra-Red (NIR) and visual (VIS) spectrum, presents challenges due to differences in illumination and sensor characteristics. NIR sensors are commonly used in practical applications, such as in nighttime outdoor surveillance systems, due to their ability to capture images in low light conditions and at various standoff distances. This study proposes a simple CSFR method that combines deep learning techniques with photometric normalization using Homomorphic Filtering (HF) to identify subjects in NIR images across different standoff distances. Both NIR and VIS images are filtered to mitigate illumination discrepancies. A deep convolutional neural network (CNN) model is employed to automatically extract discriminative deep features from the normalized images of both spectrums. Four deep CNN models of ResNet-50, DenseNet, EfficientNet, and MobileNetV2 are tested. The performance of the models in feature extraction for CSFR is evaluated on the LDHF (Long Distance Heterogeneous Face) database. The Hamming distance metric is employed to match NIR probe images with VIS gallery images. The experimental findings, expressed in recognition rate RR, demonstrate the effectiveness of the approach in enhancing CSFR accuracy, with MobileNetV2 achieving the highest accuracy of 90% for a short standoff. In addition, HF improves image quality, enhancing the deep learning model's ability to discern image characteristics.