Assessing Deep Neural Networks in Face Recognition Using Multiple Mugshot Sets

While face recognition technology using Convolutional Neural Networks (CNNs) has advanced to a level where it is widely applied in smartphone unlocking, passport verification, and even in crime investigation support systems, challenges persist in the domain of Pose-Invariant Face Recognition (PIFR). One key issue is the limited research on the effect of various mugshot orientations, beyond the standard frontal and right profile images typically captured by law enforcement, on the accuracy of face recognition in real-world security camera footage. In response, our study evaluates the performance of three CNNs for face recognition, VGG16, ResNet50, and SENet using the Face Recognition from Mugshots DataBase (FRMDB). This database is specifically designed to assess how different mugshot sets impact face recognition effectiveness. This work builds on our prior research by aligning the tests to the best practice carried out by law enforcement agencies that apply face recognition.