Explainable Face Recognition Based on Accurate Facial Compositions

Abstract

With impressive advances made in face recognition, the explainability has attracted more and more attentions in the community, which delves into traceable and well-founded clues behind the identifications in addition to the confidence scores. However, the current Explainable Face Recognition (XFR) methods are difficult to balance the explain-ability and the recognition performance. In this paper, we propose a framework based on Accurate Facial Compositions, namely AFC-XFR. The framework consists of three modules: the Backbone for feature extraction, the Local Feature Refine Module (LFRM) for semantic feature refining, and the Self-Attention based Reconstruction Module (SARM) for serialized feature interaction. Fifteen semantic features, which are accurately captured from local facial components via the proposed acquisition scheme, are conveyed in the latter two modules. Moreover, the LFRM allows us to verify three significant insights experimentally, obtaining the explainability from the perspective of model decisions. Inspired by the insight "Facial features are processed holistically", the SARM’s internal feature interaction mechanism facilitates performance increase. Extensive experiments on varying loss functions and network architectures accomplish consistent advances on evaluation benchmarks.