Scholarlink

<p>In recent years, the emergence of neural architecture search (NAS) has brought a series of breakthroughs to the field of computer vision. However, existing NAS methods either overly focus on optimizing for classification accuracy, resulting in complex networks with excessive parameters and computational requirements, or use fixed search parameters, limiting search performance. To effectively evolve neural networks, this paper proposed a lightweight search algorithm (GS-NAS). Firstly, we introduced a novel search space based on MobileBlocks and ShuffleBlocks and we employed multi-objective optimization considering classification accuracy, parameters, and FLOPs (floating-point operations per second) during the search process. Secondly, we utilized particle swarm optimization (PSO) as the search strategy, and to enhance the global search capability of the particle swarm, we integrated concepts from Grey Wolf Optimization and cosine annealing into the PSO algorithm, introducing a series of improvements to the particle update rules. Validation on standard datasets demonstrated that the improved PSO exhibited stronger global search capabilities, achieving Pareto optimality in three-objective optimization compared to various state-of-the-art algorithms. Finally, to showcase the efficiency of the obtained networks, we utilized the searched network as the backbone to design an industrial defect detection network, which achieved a leading mean Average Precision (mAP) of 79.2<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13042_2025_2686_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> on the NEU-DET dataset.</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

Evolutionary neural architecture search based on a modified particle swarm optimization

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Tao Gong,
  • Liang Wang,
  • Yongjie Ma

摘要

In recent years, the emergence of neural architecture search (NAS) has brought a series of breakthroughs to the field of computer vision. However, existing NAS methods either overly focus on optimizing for classification accuracy, resulting in complex networks with excessive parameters and computational requirements, or use fixed search parameters, limiting search performance. To effectively evolve neural networks, this paper proposed a lightweight search algorithm (GS-NAS). Firstly, we introduced a novel search space based on MobileBlocks and ShuffleBlocks and we employed multi-objective optimization considering classification accuracy, parameters, and FLOPs (floating-point operations per second) during the search process. Secondly, we utilized particle swarm optimization (PSO) as the search strategy, and to enhance the global search capability of the particle swarm, we integrated concepts from Grey Wolf Optimization and cosine annealing into the PSO algorithm, introducing a series of improvements to the particle update rules. Validation on standard datasets demonstrated that the improved PSO exhibited stronger global search capabilities, achieving Pareto optimality in three-objective optimization compared to various state-of-the-art algorithms. Finally, to showcase the efficiency of the obtained networks, we utilized the searched network as the backbone to design an industrial defect detection network, which achieved a leading mean Average Precision (mAP) of 79.2 \(\%\) % on the NEU-DET dataset.