Scholarlink

<p>Convolutional Neural Networks (CNN) optimization for specific tasks has become a critical area in research, especially in domains where accuracy and reliability are essential, such as image categorization. However, due to the large search area and computational cost, finding the optimal architecture design for a given task remains a significant challenge. The research community has experimented with a variety of optimization techniques, such as evolutionary methods and heuristic approaches, to automate CNN architecture design. This study introduces a novel hybrid optimization method, called PSO-SA, which combines Particle Swarm Optimization (PSO) with Simulated Annealing (SA) to automate CNN architecture design. The method was evaluated on multiple datasets, including RECT, MBI, and MNIST, and applied to a real-world scenario involving COVID-19 chest X-ray image classification. Results show that the PSO-SA approach, particularly the RECT_CNN model, consistently outperformed the state-of-the-art IDECNN model by margins of up to 17%, achieving up to 96.3% accuracy, 95.8% precision, 90.6% recall, and 92.8% F1-score on the COVID-19 classification task. This performance demonstrates that RECT_CNN is the most efficient architecture, outperforming the MBI_CNN and MNIST_CNN models, which achieved 96.2% and 95.2% accuracy, 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

PSO-SA: neural architecture search optimization via simulated annealing and particle swarm for image classification applications

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Manar Abu Talib,
  • Basma Alsaid,
  • Ayad Turky,
  • Qassim Nasir,
  • Takua Mokhamed

摘要

Convolutional Neural Networks (CNN) optimization for specific tasks has become a critical area in research, especially in domains where accuracy and reliability are essential, such as image categorization. However, due to the large search area and computational cost, finding the optimal architecture design for a given task remains a significant challenge. The research community has experimented with a variety of optimization techniques, such as evolutionary methods and heuristic approaches, to automate CNN architecture design. This study introduces a novel hybrid optimization method, called PSO-SA, which combines Particle Swarm Optimization (PSO) with Simulated Annealing (SA) to automate CNN architecture design. The method was evaluated on multiple datasets, including RECT, MBI, and MNIST, and applied to a real-world scenario involving COVID-19 chest X-ray image classification. Results show that the PSO-SA approach, particularly the RECT_CNN model, consistently outperformed the state-of-the-art IDECNN model by margins of up to 17%, achieving up to 96.3% accuracy, 95.8% precision, 90.6% recall, and 92.8% F1-score on the COVID-19 classification task. This performance demonstrates that RECT_CNN is the most efficient architecture, outperforming the MBI_CNN and MNIST_CNN models, which achieved 96.2% and 95.2% accuracy, respectively.