Adjusted Attention YOLOX-Based Far-Distance Face-Recognition
Chih‐Lyang Hwang, Zih‐En Cheng · 2023
To satisfy the required recognition accuracy from a far distance, an adjusted attention-based YOLOX for face recognition (AA-YOLOX-FR) is designed by an appropriate segmentation of the original image, so that faces' pixels (e.g., 20 × 20 at 15m) effectively train, validate, and test. Based on an edge computing platform (e.g., NVIDIA Jetson-AGX), the processing time for the image 2208 × 1242 with 8 segmentations of 640 × 640 equals 296.3ms in comparison to 95.4ms for its down-sampling to 640 × 640. Although the down-sampling technique achieves a faster processing speed, its recognition rate decreases as a face is at a far distance or with different lighting conditions. The average online video-based recognition rate of AA-YOLOX-FR for a distance from 10m to 15 m and different lighting conditions is 92.5%.