An Efficient Facial Verification System for Surveillance that Automatically Selects a Lightweight CNN Method and Utilizes Super-Resolution Images

In the last decade, facial recognition and verification methods have been extensively used for surveillance and security purposes. However, most of the time, recognizing and/or verifying faces is challenging due to the low facial resolution of the obtained or captured images. Likewise, low-resolution facial images contain different facial features, such as variations in pose, lighting, resolution, and camera-to-subject distance. Moreover, the methods commonly used for facial verification in images are based on deep architectures, which rely on complex deep learning models with promising verification results but come with high computational costs. On the other hand, real-world requirements for facial verification demand lightweight methods that are inspired by their counterparts but are also compact and efficient enough to be used in unrestricted scenarios, such as video surveillance or security cameras. This paper proposes a lightweight facial verification system (LFVS) that automatically selects a lightweight facial verification method based on images characteristics, such as facial rotation variations and low resolutions. Additionally, a dynamic scaling approach is proposed to upscale images to the required size. The system uses this scaling to enhance the image resolution and succeeded in improving facial verification performance. Experimental results demonstrated that the lightweight facial verification system achieved better results when using super-resolution images in low-resolution reference points, with minimal memory usage and computational complexity.