D2SC-GAN: Dual Deep-Shallow Channeled Generative Adversarial Network, for Resolving Low-Resolution Faces for Recognition in Classroom Scenarios

Abstract

Face Recognition using convolutional neural networks have achieved considerable success in constrained environments in the recent past. However, the performance of these methods deteriorates in case of mismatch of training and test distributions, under classroom/surveillance scenarios. These test (probe) samples suffer from degradations such as noise, poor illumination, pose variations, occlusion, low-resolution (LR), blur as well as aliasing, when compared to the crisp, rich training (gallery) set, comprising mostly of high-resolution (HR) mugshot images captured in laboratory settings. To cope with this scenario, we propose a novel dual deep-shallow channeled generative adversarial network (D2SC-GAN) which performs supervised domain adaptation (DA) by mapping LR degraded probe samples to their corresponding HR gallery-like counterparts to perform closed-set face recognition. D2SC-GAN uses a multi-component loss function comprising of multi-resolution patchwise MSE and normalized chi-squared distance loss functions, along with a Kullback-Leibler divergence based loss function. Moreover, we propose a novel classroom face dataset called the Indian Classroom Face Dataset (ICFD), which, to the best of our knowledge, is a first of its kind and will be helpful to explore the challenges of face recognition when used for automatically recording the attendance in classroom conditions. The proposed network achieves superior results on five real-world face datasets when compared with recent state-of-the-art deep as well as shallow supervised domain adaptation (DA), super-resolution (SR), and degraded face recognition (DFR) methods, which show the effectiveness of our proposed method.