Deep Learning Methods for Homogeneous and Heterogeneous Face Recognition

Abstract

Visible face recognition systems are subjected to failure when recognizing the faces in unconstrained scenarios. So, recognizing faces under variable and low illumination conditions are more important since most of the security breaches happen during night time. Near Infrared (NIR) spectrum enables to acquire high quality images, even with- out any external source of light and hence it is a good method for solving the problem of illumination. The proposed method addresses three different problems namely: face recognition, gender classification and facial expression recognition in the homogeneous NIR spectrum. In order to perform homogeneous Near Infrared (NIR) face recognition two methods have been proposed to recognize face in NIR spectrum. First method is based on transfer learning for face feature extraction and classification using three, sep- arate SVM classifiers, which is a modified ResNet-34 model with SVM classifier. In contrast, to reduce the computational overhead further an end-to-end light CNN model has also been proposed to perform the above mentioned tasks. The methods have been analyzed on the publicly available, challenging, benchmark datasets CASIA NIR-VIS 2.0, Oulu-CASIA NIR-VIS, PolyU, CBSR, IIT Kh and HITSZ for face recognition. It has been observed that the Light CNN model has given high face recognition accuracies of 98.22%, 97.90%, 98.16% and 99.26% for CBSR, PolyU, CASIA NIR-VIS 2.0 and Oulu-CASIA datasets respectively with limited number of trainable parameters (0.285M). Further, in the case of gender classification, the ResNet-34 model with SVM classifier has given high accuracies and it has been observed to be 95.37%, 98.82%, 98.93% and 99.44% for CBSR, PolyU, IIT Kh and Oulu-CASIA datasets respectively. Finally, the facial expression recognition accuracies are 81.24% and 71.36% on Oulu- ii CASIA dataset respectively for the Light CNN model (proposed method 2) which is comparatively high when compared to the results obtained for the modified ResNet-34 model with SVM (pr