Improving Real-Time Attendance System Based On Yolo and Facenet Using Super Resolution Models

A real-time attendance system that uses the face recognition model can reduce workload rather than manual attendance. One of the problems of the real-time attendance system with face recognition is the low quality of face images detected from a long distance. The solution To overcome this problem is by combining several deep learning models like YOLOv11, super-resolution (SR) models that use GAN architecture like ESRGAN, Real-ESRGAN, and GFPGAN, Facenet is trained with Casia-WebFace and VGG-Face2 datasets. All datasets and detected faces are aligned with Mediapipe. Face recognition is done by measuring the distance between embeddings using cosine similarity. The dataset contains Indonesian faces with various poses, expressions, and illuminations. The dataset is processed again to become two datasets, one resized to 64x64 pixels and one not resized. Tests were carried out using video recordings containing the activities of students who entered the classroom simultaneously and were taken using a webcam remotely. The results show that Facenet trained with VGG-Face2 is very suitable for the Indonesian dataset. GFPGAN can produce more realistic facial images than other SR models. The resized dataset can improve the quality of face recognition because the resized dataset is suitable for detected faces that have some missing facial information. YOLOv11m-Face and YOLOv11n-Face can improve the quality of face recognition than combination of YOLOv11s-Face and GFPGAN there are unrecognized faces. The test results recommended a combination of models (YOLOv11m-Face or YOLOv11n-Face), GFPGAN, Facenet is trained with VGG-Face2 dataset, and resized dataset.