NAS For efficient mobile eyebrow biometrics

Occlusions, such as those due to wearing surgical masks, pose a significant challenge to the face recognition systems. Among possible remedies, ocular biometric has proven to be a popular choice. However, the upper ocular regions, especially the patterns presented by the eyebrows, have yet to gain the attention they deserve. In this work, we leverage Neural Architecture Search (NAS) to discover better-performing architectures for eyebrow recognition. To reduce the computational complexity, we apply a zero-shot NAS to assess the exploratory architectures’ performance prior to any training. We were able to discover three new architectures that achieved competitive accuracies in eyebrow recognition. In doing so, we explored depthwise separable convolution, hard-swish, and Arcface loss functions to further enhance the discovered models in terms of accuracy and number of parameters. Our best result provided 0.999 AUC, 0.6% EER, and 98.25% GMR at FMR over FACES dataset, which is better than the results of state-of-the-art architecture, a 29-layer lightCNN which has 21 more parameters and 8 more FLOPS.