Scholarlink

<p>Masked face recognition presents unique challenges especially due to occlusion caused by face masks. This paper introduces a novel approach called HOGE which integrates the visualisation of HOG images as input for a modified EfficientNetV2-S model to perform masked face recognition. This modified EfficientNetV2-S model employs an additional convolution layer for processing greyscale masked face images which differs from other pre-trained CNN-based models. It aims to address the challenge of recognising masked face images and investigate the effect of feature descriptor interaction with deep transfer learning technique in masked face recognition. Two benchmark datasets were used to evaluate the performance of proposed method. Experimental results demonstrate that HOGE achieves accuracies of 97.41% on LFW-SMFRD and 99.38% on RMFRD which show that the proposed method works well and effectively recognises masked face images from both datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HOGE: integrating feature descriptor and transfer learning for masked face recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月15日

查看全文
  • Ming Chun Yo,
  • Siew Chin Chong,
  • Lee Ying Chong

摘要

Masked face recognition presents unique challenges especially due to occlusion caused by face masks. This paper introduces a novel approach called HOGE which integrates the visualisation of HOG images as input for a modified EfficientNetV2-S model to perform masked face recognition. This modified EfficientNetV2-S model employs an additional convolution layer for processing greyscale masked face images which differs from other pre-trained CNN-based models. It aims to address the challenge of recognising masked face images and investigate the effect of feature descriptor interaction with deep transfer learning technique in masked face recognition. Two benchmark datasets were used to evaluate the performance of proposed method. Experimental results demonstrate that HOGE achieves accuracies of 97.41% on LFW-SMFRD and 99.38% on RMFRD which show that the proposed method works well and effectively recognises masked face images from both datasets.