ROI-YOLOv8-Based Far-Distance Face-Recognition
Felix Gunawan, Chih‐Lyang Hwang, Zih‐En Cheng
Abstract
This paper presents a model for far-distance face recognition using ROI-YOLOv8. We achieve this by training YOLOv8 for 3 target faces on our custom datasets: (i) The first dataset is the original dataset with training and validation images of 2412 and 229, respectively. (ii) The second dataset augments the more pixelated images from the 1 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">st</sup> one with training and validation images of 4824 and 458, respectively. (iii) The third dataset considers various exposure, noise, and blur from the 2 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> one with training and validation images of 14,272 and 459, respectively. To enhance the far-distance recognition, a two-stage recognition is considered. At first, a pre-trained YOLOv8 model for human detection is achieved. A Region-of-Interest (ROI) including detected humans is segmented as the size of 640 × 640 pixels for the input of another YOLOv8, i.e., ROI-YOLOv8-FR. A computer with Intel i5-12400F, 16GB RAM, and NVIDIA RTX 3080Ti with 12GB VRAM is used as computing platform. The trained times for these 3 datasets are 50, 95, and 219 minutes, respectively. Their mAP50s are 99.5% and mAP50-95s are slightly different as 88.112%, 87.962%, and 88.103% respectively. More important, the 3 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rd</sup> trained model can successfully recognize a face at 30 and 35m with the confidences of 65.1% and 75.6%.