Scholarlink

<p>Developing accurate facial recognition (FR) methods using Single Sample Per Person (SSPP) remains a significant challenge. While Advanced Deep Learning (ADL) techniques have improved FR accuracy, they typically require extensive, well-labeled datasets captured from multiple angles. This results in high data collection costs, increased labeling errors, and reliance on preprocessing techniques such as noise removal and image enhancement. To address these challenges, we propose a novel framework that optimally integrates established methods to enhance FR efficiency and performance without requiring multiple images per person (MIPP) or image enhancement techniques. Our approach consists of three key stages: (1) the Viola-Jones method for rapid face detection, (2) a SubClass Convolutional Neural Network (SC-CNN) for feature extraction and encoding via serialization, and (3) a pattern-matching technique for classification. The novelty lies in the strategic combination and optimization of these components, enabling robust SSPP-based FR while meeting Quality of Service (QoS) criteria, such as reducing the number of iterations and minimizing makespan, without compromising accuracy. Experimental results show that our framework outperforms traditional methods—including S-CNN, RNN, MLP, and DBN-achieving accuracy improvements of 7.25%, 8.42%, 3.73%, and 2.45%, respectively.</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 multi-stage deep learning approach for facial recognition and classification

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Chirag Chandrashekar,
  • Maheswari Subburaj,
  • Arun Kumar Sivaraman,
  • Ummity Srinivasa Rao,
  • Janakiraman Nithiyanantham,
  • Ajmery Sultana

摘要

Developing accurate facial recognition (FR) methods using Single Sample Per Person (SSPP) remains a significant challenge. While Advanced Deep Learning (ADL) techniques have improved FR accuracy, they typically require extensive, well-labeled datasets captured from multiple angles. This results in high data collection costs, increased labeling errors, and reliance on preprocessing techniques such as noise removal and image enhancement. To address these challenges, we propose a novel framework that optimally integrates established methods to enhance FR efficiency and performance without requiring multiple images per person (MIPP) or image enhancement techniques. Our approach consists of three key stages: (1) the Viola-Jones method for rapid face detection, (2) a SubClass Convolutional Neural Network (SC-CNN) for feature extraction and encoding via serialization, and (3) a pattern-matching technique for classification. The novelty lies in the strategic combination and optimization of these components, enabling robust SSPP-based FR while meeting Quality of Service (QoS) criteria, such as reducing the number of iterations and minimizing makespan, without compromising accuracy. Experimental results show that our framework outperforms traditional methods—including S-CNN, RNN, MLP, and DBN-achieving accuracy improvements of 7.25%, 8.42%, 3.73%, and 2.45%, respectively.