Scholarlink

<p>This research introduces the Enhanced Face Classification in Dynamic Scenarios - Pose, Lighting, and Occlusion (EFCDS-PLO) framework, designed to tackle the critical challenges of rapid pose variations, inconsistent lighting, and frequent partial occlusions in dynamic video contexts, which significantly complicate face recognition compared to controlled environments. Using the sitcom Friends Season 1 dataset (1,147 frames featuring 6 main characters with partial occlusions, diverse camera angles, and bright, moderate, low-light scenarios) as a representative testbed, the framework integrates multi-scale feature fusion, adaptive histogram equalization, quality-aware filtering, and margin-expanded cropping to enhance robustness. Spatio-temporal trajectory modeling with weighted voting further optimizes classification for dynamic interactions. Experimental results show that EFCDS-PLO achieves a maximum mAP of 0.9965 with MobileNet-SSD and an F1-score of 0.8164 with YOLO, outperforming baselines. Ablation studies validate that quality filtering and margin cropping significantly improve resilience to real-world variations. This work advances dynamic face recognition for complex video scenarios like sitcoms, providing a robust solution for unconstrained environments where traditional methods struggle with pose, lighting, and occlusion challenges.</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

Robust face classification under pose, lighting, and occlusion challenges: friends as testbed

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Gao Yousheng,
  • Raseeda Hamzah,
  • Li Ang,
  • Nurbaity Sabri

摘要

This research introduces the Enhanced Face Classification in Dynamic Scenarios - Pose, Lighting, and Occlusion (EFCDS-PLO) framework, designed to tackle the critical challenges of rapid pose variations, inconsistent lighting, and frequent partial occlusions in dynamic video contexts, which significantly complicate face recognition compared to controlled environments. Using the sitcom Friends Season 1 dataset (1,147 frames featuring 6 main characters with partial occlusions, diverse camera angles, and bright, moderate, low-light scenarios) as a representative testbed, the framework integrates multi-scale feature fusion, adaptive histogram equalization, quality-aware filtering, and margin-expanded cropping to enhance robustness. Spatio-temporal trajectory modeling with weighted voting further optimizes classification for dynamic interactions. Experimental results show that EFCDS-PLO achieves a maximum mAP of 0.9965 with MobileNet-SSD and an F1-score of 0.8164 with YOLO, outperforming baselines. Ablation studies validate that quality filtering and margin cropping significantly improve resilience to real-world variations. This work advances dynamic face recognition for complex video scenarios like sitcoms, providing a robust solution for unconstrained environments where traditional methods struggle with pose, lighting, and occlusion challenges.