CNN Based Periocular Recognition Using Multispectral Images

Abstract

Over the recent years, the periocular region has emerged as a potential unconstrained biometric trait for person authentication. For a biometric identification scenario to operate reliably round the clock, it should be capable of subject recognition in multiple spectra. However, there is limited research associated with the non-ideal multispectral imaging of the periocular trait. This is critical for real life applications such as surveillance and watch list identification. The existing techniques for multispectral periocular recognition rely on fusion at the feature level. However, these handcrafted features are not primarily data driven and there even exists possibilities for more novel features that could better describe the same. One possible solution to address such issues is to resort to the data driven deep learning strategies. Accordingly, we propose to apply the attributes extracted from pretrained CNN for subject authentication. To the best of our knowledge, this is the first study of multispectral periocular recognition employing deep learning. For our work, the IIITD Multispectral Periocular (IMP) database is used. The best classification accuracy reported for this dataset is 91.8%. This value is not precise enough for biometric identification tasks. The off-the-shelf CNN features employed in our work gives an improved accuracy of 97.14% for the multispectral periocular images.