Exploring Image Augmentation Methods for Long-Distance Face Recognition Using Deep Learning

Developing an efficient long-range face recognition (FR) system involves multiple challenges that impact image quality and, as a result, FR performance. Such challenges include camera quality and settings, atmospheric conditions, non-cooperative subjects, face pose, and unfavorable lighting. To improve face image quality under such conditions, face restoration models have been proposed, which require using large-scale face datasets and augmentation methods to create low-quality, high-quality face image pairs for the restoration models to train on. However, choosing these augmentation methods is complex and can result in significant fluctuations in face recognition (verification or identification) performance. In this work, we explore the utilization of various image augmentation methods to generate pairs of low-quality images. These pairs are intended for training deep face restoration models, which will be integrated into an end-to-end long-range FR system. We assess our method's performance against benchmarks, achieving significant improvements, namely a 5% increase in Rank-1 accuracy, a 9% increase in Rank-5 accuracy, a 5% increase in AVC, and a 50% reduction in EER. These enhancements are achieved by employing Defocus Blur as the primary augmentation method for GAN Prior Embedded Network (GPEN). The dataset used for this work is a subset of the original MILAB-VTF(B) dataset, which includes indoor, high-quality face images of enrolled subjects that are matched against their outdoor 300-meter (~984ft) face image low-quality counterparts. This subset of faces simulates scenarios for long-range FR applications, such as perimeter security at airstrips, security at open-desert military bases, and similar environments. In these scenarios, subjects may be enrolled in the database during indoor sessions and subsequently matched against long-distance data.