Harnessing NNIA for Intelligent Triplet Generation in Fine-Tuned SENet-50 Based Face Recognition

Abstract

FaceNet, founded on a triplet loss function, demonstrates strong face recognition performance. However, its effectiveness hinges significantly on the triplet selection strategy, which often fails to identify the most informative triplets, especially on largescale or complex datasets. To address this, we propose a training framework incorporating Multi-Objective Evolutionary Algorithms (MOEAs) into the triplet selection process. Our approach provides intelligent selection of triplets through a bi-objective optimization function that maximizes the anchor-negative distance (to enforce inter-class separability) while simultaneously optimizing the triplet-margin constraint (to facilitate active selection of hard and semi-hard negatives). In particular, we investigate the use of Fast Non-Dominated Sorting (FNDS), a Non-dominated Neighbor Immune Algorithm (NNIA), and a modified multi objective artificial bee colony algorithm (MOABC) for such selection. This offline, optimization-based approach incurs a large computational cost for triplet selection, but produces a filtered pool of high-quality triplets, enabling a more efficient learning of highly discriminative facial features. Extensive experiments on a range of benchmark datasets—including LFW, YTF, IJB-B, AgeDB, CFP-FP, CALFW, CPLFW, as well as masked and unmasked faces of the LFW dataset—demonstrate improvements over state-of-the-art methods in both recognition accuracy and training convergence.

DOI: 10.61416/ceai.v27i3.9462