An Empirical Evaluation of Deep Architectures on Generalization of Smartphone-based Face Image Quality Assessment

Often biometric authentication relies on the quality of enrolment and probe sample and it is therefore essential to estimate the image quality before a sample is submitted to the enrolment or verification process. The challenges encountered in estimating the quality is due to generalizability over unknown data samples of different origin. To this extent, we try to evaluate various deep learning networks which in theory, show high-performance in generalization. Due to the massive adoption of biometrics in consumer solutions like smartphones, we have chosen Smartphone based Face Recognition Systems (FRS) to carry out our study. The main factors which impact the operating performance of the FRS are illumination, pose, occlusions and facial expressions. Therefore, it is essential to understand and estimate the quality of a facial image accurately. In this paper we present a robust and accurate quality estimating framework using deep neural networks (DNN). This work leverages the benefits of deep learning by transferring the pre-learned features from already trained DNNs such as AlexNet and Inception to estimate the facial image quality. Furthermore, we present the evaluation results for more than 10 techniques and 5 face image databases to analyze the performance generalization, and our results favor the pre-trained DNN models over the hand-crafted methods.