Study on Improved YOLOv10 Face Recognition Based on WF-EMA

In this paper, we propose an improved YOLOv10 model focusing on forged face recognition. Based on YOLOv10 released by Tsinghua University 2024, the model incorporates filters before the input to highlight the features and introduces an efficient multi-scale attention module (EMA), which aims to improve the feature representation capability and reduce the noise interference. Experimental results show that the improved model achieves 89.5%, 87.2%, 92.3% and 91.0% in precision, recall, average accuracy and average confidence, respectively, which significantly outperforms the baseline model's performance. Meanwhile, the study explores the effects of different modules on the model performance and verifies the effectiveness of the optimization strategy. These improvements provide a more accurate and reliable solution for forged face recognition, which helps to meet the increasing challenges of forging technology.