One-shot cross-dataset palmprint recognition via adversarial domain adaptation

Abstract

Deep learning-based palmprint recognition algorithms have obtained promising performance. However, the previous methods require a large amount of labeled samples, which are difficult to obtain. In this paper, a novel cross-dataset palmprint recognition method is proposed using as low as one labelled sample per subject in the target palmprint dataset based on adversarial domain adaptation. Two different palmprint datasets are adopted as source dataset and target dataset. The training samples from two datasets are grouped into four categories. MobileFaceNets-based deep hashing network (DHN) is introduced to extract discriminative features, which can improve the efficiencies of feature extracting and matching. To align the features in two datasets, a typical adversarial discriminator is augmented to distinguish between the four different categories. With adversarial learning, the target network is becoming adaptive to the unlabeled target palmprint images. Extensive experiments on the benchmarks including constrained and unconstrained palmprint databases demonstrate that our method can outperform the baseline models on cross-dataset palmprint verification and identification.