Interpretability Study of Face Features Within a Framework for Automatic Face Recognition

With the rapid development of face recognition technology based on deep learning, the technology has been widely used in the fields of automatic driving and security monitoring. However, it is limited to the face features based on deep learning are difficult to interpret, which hinders the deep combination of the same identification of forensic science portraits, face 1 to N comparison and artificial intelligence, and affects the application of automatic face recognition in forensic science to a certain extent. In this regard, around the task of interpretability of face features, based on CASIA-FaceV5,LFW,MTFL and other public datasets, combined with the dlib face detection algorithm, to construct an interpretable two-tuple dataset for face recognition with the main focus on redrawn portraits. The number of redrawn portraits is expanded by more than one order of magnitude (10 times) compared with the number of original portraits, and the redrawn region focuses on the key features of portraits. On the basis of integrating two different types of saliency mapping algorithms, combining the mask generation strategy and multi-scale feature fusion, we optimize the way of generating the saliency map of face features. Combined with the two-tuple dataset, the accuracy of the deep learning-based face matching system is initially quantitatively analyzed, so as to improve the application of face matching methods based on face feature similarity scores in the field of scientific image evidence in court.