Low-Resolution Face Recognition in Multi-person Indoor Environments Using Convolutional Neural Networks

Face recognition has been widely applied in many systems in our lives. These applications have reached good accuracies on face recognition tasks when face images can be captured with good quality, particularly when they have a high enough resolution. However, in an indoor environment, a surveillance camera often covers a wide area with multiple persons; this leads to only lower resolutions of face images are available in face recognition. This paper presents a face recognition approach for low resolution images using convolutional neural network (CNN) in a multi-person indoor environment. Our methods first detect face regions with the YOLOv3 approach and then recognize face images using the trained CNN model. Experiments are performed in an indoor classroom to capture face images with resolutions ranging from 20x20 to 70x70. Moreover, face images are extracted over 4 months to test the stability of our proposed face recognition model.