Face recognition is usually used in terms of security, even more so in terms of surveillance technology. Cross spectral face recognition is a face recognition system that uses different spectrums. For example, the image is taken using a camera with a VIS spectrum and then matched with an image taken using a camera with an NIR spectrum. Due to the different lighting conditions when taking the image, the two image results are also different. This difference in lighting conditions is a challenge in cross spectral face recognition. This study implements the Convolutional Neural Network (CNN) method using the MobileNetV2 architecture for facial recognition in the Long-Distance Heterogeneous Face database (LDHF-DB) dataset. This study tested three different scenarios, namely at a distance of 1m, 60m, and 100m. The accuracy values obtained by applying homomorphic filtering and without homomorphic filtering at a distance of 1m are 90% and 73%, then at a distance of 60m the accuracy values are 61% and 60%, and at a distance of 100m the accuracy values are 50% and 49%. . Based on trials, the use of homomorphic filtering in the preprocessing stage can increase accuracy up to 17%. This study proves that in cross spectral image face recognition, the MobileNetV2 architecture is very well used so that it can produce 90% accuracy.
Keywords : Face Recognition, cross spectral, MobileNetV2