Face Recognition Method for Underground Engineering Based on Dual-Target Domain Adaptation and Discriminative Feature Learning

The poor lighting and dust in underground engineering environments often induce serious noise and motion blur to face recognition cameras, decreasing the recognition accuracy even under Near-InfraRed (NIR) mode. Although various methods to improve algorithm robustness can enhance the overall model performance, the face recognition accuracy remains insufficient for the large number of blurred and NIR images generated during underground engineering. Accordingly, this paper proposes a novel face recognition method for underground engineering environments based on dual-target domain adaptation (DTDA) and discriminative feature learning (DFL). The targeting of the algorithm to blurred and NIR images is enhanced through DTDA, while the inter-class feature variability and intra-class feature compactness for domain adaptation face recognition are ensured by DFL. Extensive experiments are conducted on public and self-built datasets to verify the proposed method. The results show that (1) dual domain adversarial loss, dual CORAL loss, Blur-to-Sharp and NIR-to-VIS image reconstruction, and face keypoints intermediate domain can improve the DTDA performance of the model. (2) Dedicated encoders, center loss, and triplet loss can improve the DFL ability of the model. (3) The high Rank-1 accuracies and area under curve (AUC) values of the Receiver Operating Characteristic (ROC) curves on the three domains indicate that the proposed face recognition method can adapt to the complex working conditions in underground engineering environments.