Abstract
Nowadays, it is imperative to accurately detect and repair face flaws to obtain high-resolution, detail-preserving face images in social life. However, there exists a significant gap between face flaws and natural scene features. In this paper, we propose a framework of HRT-YOLO to erase the problems of small scale, irregular shape, and regional overlap of face flaws. Based on YOLOv5, we introduce high-resolution representation, where the main network of high-resolution feature maps is gradually added in parallel to the sub-network of low-resolution feature maps, allowing for more accurate localization of flaws. We also implement transformer detection heads, which use a multi-head attention to mine contextual features and improve the defect-intensive regions detection. Finally, we present a high-resolution face flaw dataset to evaluate our approach. The evaluation results show that HRT-YOLO has excellent performance on face flaw detection, with 3.23% improvement in mAP and 5.33% improvement in accuracy compared to the baseline (YOLOv5).