2021/11/16 by Domenick Poster, Poster, Domenick, Nasser M. Nasrabadi +1
Medicine · #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #Cutaneous Melanoma Detection and Management #FOS: Computer and information sciences #Facial Rejuvenation and Surgery Techniques #Nasal Surgery and Airway Studies
paper · pdf · doi:10.48550/arxiv.2111.08738
openalex publication_date 2021/11/16 · openalex created_date 2022/10/24 · openalex updated_date 2026/07/28
A common yet challenging scenario in periocular biometrics is cross-spectral\nmatching - in particular, the matching of visible wavelength against\nnear-infrared (NIR) periocular images. We propose a novel approach to\ncross-spectral periocular verification that primarily focuses on learning a\nmapping from visible and NIR periocular images to a shared latent\nrepresentational subspace, and supports this effort by simultaneously learning\nintra-spectral image reconstruction. We show the auxiliary image reconstruction\ntask (and in particular the reconstruction of high-level, semantic features)\nresults in learning a more discriminative, domain-invariant subspace compared\nto the baseline while incurring no additional computational or memory costs at\ntest-time. The proposed Coupled Conditional Generative Adversarial Network\n(CoGAN) architecture uses paired generator networks (one operating on visible\nimages and the other on NIR) composed of U-Nets with ResNet-18 encoders trained\nfor feature learning via contrastive loss and for intra-spectral image\nreconstruction with adversarial, pixel-based, and perceptual reconstruction\nlosses. Moreover, the proposed CoGAN model beats the current state-of-art\n(SotA) in cross-spectral periocular recognition. On the Hong Kong PolyU\nbenchmark dataset, we achieve 98.65% AUC and 5.14% EER compared to the SotA EER\nof 8.02%. On the Cross-Eyed dataset, we achieve 99.31% AUC and 3.99% EER versus\nSotA EER of 4.39%.\n