Scholarlink

<p>Occlusion is a significant challenge in face recognition accuracy. An innovative occlusion-aware approach is presented in this paper to mitigate the impact of occlusions on recognition performance. Unlike traditional methods that require extensive occlusion data, our approach utilizes a general occlusion perception network with facial semantic parsing. This network classifies “non-face”regions as occlusions, which makes our approach robust against various occlusion types. For the precise localization of occluded areas, facial contour detection is employed, and missing contour segments are predicted, with this contour information being utilized in subsequent recognition processes. In the recognition phase, the face is segmented into sub-regions based on spatial and semantic cues. A network, similar to Pairwise Difference Siamese Network, is developed to construct a region-feature dictionary, which enables the selective suppression of features affected by occlusion. The efficacy of the proposed method is demonstrated through extensive experiments conducted on the AR, Megaface, LFW, and synthetic occlusion 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

Research on occlusion perception facial feature correlation based on less-class learning

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Yaqian Li,
  • Xin Zhan,
  • Ya Gao,
  • Haibin Li,
  • Wenming Zhang

摘要

Occlusion is a significant challenge in face recognition accuracy. An innovative occlusion-aware approach is presented in this paper to mitigate the impact of occlusions on recognition performance. Unlike traditional methods that require extensive occlusion data, our approach utilizes a general occlusion perception network with facial semantic parsing. This network classifies “non-face”regions as occlusions, which makes our approach robust against various occlusion types. For the precise localization of occluded areas, facial contour detection is employed, and missing contour segments are predicted, with this contour information being utilized in subsequent recognition processes. In the recognition phase, the face is segmented into sub-regions based on spatial and semantic cues. A network, similar to Pairwise Difference Siamese Network, is developed to construct a region-feature dictionary, which enables the selective suppression of features affected by occlusion. The efficacy of the proposed method is demonstrated through extensive experiments conducted on the AR, Megaface, LFW, and synthetic occlusion datasets.