Deep learning features in facial identification and the likelihood ratio bound

Fine alignment, flexible fusion: A novel framework of multi-model score fusion in face identification

Multi-model score fusion was considered a bottleneck problem in forensic face identification. While the score distribution of different face models varies greatly, the existing score processing methods cannot achieve accurate alignment. This paper proposed a score fusion framework named fine alignment and flexible fusion framework (FAFF). In FAFF, we took score-based likelihood ratios as the reference values to align the similarity scores generated by different face models. First, we set up a unified calibration test workflow based on the forensic likelihood ratio test. Then, 3 LLR anchor-based methods (LLRBA1, LLRBA2, and LLRBA3) and LLR curve-based methods (LLRBC) were proposed. Finally, we conducted fusion experiments on four face models (VGGface, Facenet, Arcface, and SFace). The experimental results show that on the CelebA dataset, compared with the existing MOEBA and PAN methods, LLRBC increased the TPR@ 10-7 FPR by 175.4 % and 162.9 %, and LLRBA increased by 55.6 % and 48.5 %.

Two heads are better than one: Dual systems obtain better performance in facial comparison

Forensic facial image comparison based on recognition algorithms has been widely applied in forensic science. Previous researches have been concentrating on the cases of using single system during comparison, while how to use multiple systems has not yet been studied. In this paper, a dual-systems model (including SeetaFace and FaceNet) for facial comparison was constructed, and Bayesian networks were utilized as the basic frame. In order to prove its superiority, a large-scale experiment (on the dataset CelebA) has been carried on to evaluate the score-based likelihood ratio. We used three likelihood ratio evaluation tools (Empirical Cross-Entropy, Cost Likelihood Ratio, Limit Tippett Plots) to assess the performance of the model. The Wasserstein distance was also used to evaluate the detailed likelihood ratio performance. The experimental results show that the likelihood ratio performance of our dual-systems model is better than single system. Besides, our method of model building and evaluation can also be used in the condition of triple or more systems.