Scholarlink

<p>Deep convolutional neural networks (CNNs) are powerful deep learning models for tackling various computer vision challenges, particularly image classification. However, designing an optimal CNN model is a complex, labor-intensive, and resource-demanding process that typically requires extensive manual effort and domain-specific knowledge. This paper introduces EPSO-CNN, an efficient framework leveraging particle swarm optimization (PSO) to automatically design optimal CNN architectures for image classification tasks. The proposed method utilizes a modified variable-length encoding scheme for flexible and efficient representation of CNN layers as particles, facilitating rapid architecture construction without conversion overhead. Novel strategies for velocity and position updates are incorporated to enhance search exploration and prevent premature convergence. Batch normalization and dropout techniques are employed to accelerate training and mitigate overfitting. Experimental results on eight widely used image classification datasets show that EPSO-CNN outperforms most of the 26 state-of-the-art models with notably lower error rates, and its superiority is further validated on NAS-Bench-101 and NAS-Bench-201, confirming its competitiveness across standardized NAS evaluation platforms. The best CNN architectures generated by EPSO-CNN are evaluated on facial expression recognition (FER) challenge dataset, showcasing their robustness and generalization ability in real-life application. The findings confirm the efficacy of EPSO-CNN in automating the design of deep CNN architectures, offering a valuable tool for advancing image classification methodologies.</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

Automated CNN architecture design with enhanced particle swarm optimization

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Arjun Ghosh,
  • Nanda Dulal Jana,
  • Subhayu Ghosh

摘要

Deep convolutional neural networks (CNNs) are powerful deep learning models for tackling various computer vision challenges, particularly image classification. However, designing an optimal CNN model is a complex, labor-intensive, and resource-demanding process that typically requires extensive manual effort and domain-specific knowledge. This paper introduces EPSO-CNN, an efficient framework leveraging particle swarm optimization (PSO) to automatically design optimal CNN architectures for image classification tasks. The proposed method utilizes a modified variable-length encoding scheme for flexible and efficient representation of CNN layers as particles, facilitating rapid architecture construction without conversion overhead. Novel strategies for velocity and position updates are incorporated to enhance search exploration and prevent premature convergence. Batch normalization and dropout techniques are employed to accelerate training and mitigate overfitting. Experimental results on eight widely used image classification datasets show that EPSO-CNN outperforms most of the 26 state-of-the-art models with notably lower error rates, and its superiority is further validated on NAS-Bench-101 and NAS-Bench-201, confirming its competitiveness across standardized NAS evaluation platforms. The best CNN architectures generated by EPSO-CNN are evaluated on facial expression recognition (FER) challenge dataset, showcasing their robustness and generalization ability in real-life application. The findings confirm the efficacy of EPSO-CNN in automating the design of deep CNN architectures, offering a valuable tool for advancing image classification methodologies.