Scholarlink

<p>Driven by advancements in industrial big data, automatic metal surface defect detection has been granted unprecedented attention in industrial product quality control. However, models are required to extract both global and local features simultaneously due to the diversity and complexity of metal surface defects, which are further complicated by image quality issues. Consequently, a classification model utilizing hierarchical multi-scale feature fusion (HMFF) is proposed. The model is structured with a parallel architecture that integrates Swin Transformers and ConvNeXt, enabling the extraction of global and local features across various semantic scales. An adaptive hierarchical feature fusion layer (AHFF) is designed to optimize the utilization and fusion of features across different scales, which integrates channel attention (CA), spatial attention (SA), and a multi-layer inverted residual block (ML-IRB) for adaptive semantic information fusion. Specifically, inter-channel dependencies are addressed by the CA module, while local features are amplified by the SA module. To further enhance the efficacy of feature fusion, the ML-IRB, innovatively embedded within this layer, is utilized to capture global and local feature information across multiple levels. Experimental results indicate that an accuracy rate of 99.6% is achieved by the HMFF model on the NEU-DET dataset, and 96.9% on the GC10-DET dataset, reflecting improvements of 2.1, 3.8, 3.1, and 2.2% over the ConvNeXt, Swin Transformer, VGG16, and ResNet34 models, respectively, and 4.8, 6.0, 2.6, and 2.8% over other advanced models. The HMFF model outperforms the other models in terms of performance.</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

A Hierarchical Multi-Scale Feature Fusion-Based Classification Model for Metal Defects

  • 期刊论文

  • 发表日期：

    2025年5月22日

查看全文
  • Jitong Li,
  • Jie Liu

摘要

Driven by advancements in industrial big data, automatic metal surface defect detection has been granted unprecedented attention in industrial product quality control. However, models are required to extract both global and local features simultaneously due to the diversity and complexity of metal surface defects, which are further complicated by image quality issues. Consequently, a classification model utilizing hierarchical multi-scale feature fusion (HMFF) is proposed. The model is structured with a parallel architecture that integrates Swin Transformers and ConvNeXt, enabling the extraction of global and local features across various semantic scales. An adaptive hierarchical feature fusion layer (AHFF) is designed to optimize the utilization and fusion of features across different scales, which integrates channel attention (CA), spatial attention (SA), and a multi-layer inverted residual block (ML-IRB) for adaptive semantic information fusion. Specifically, inter-channel dependencies are addressed by the CA module, while local features are amplified by the SA module. To further enhance the efficacy of feature fusion, the ML-IRB, innovatively embedded within this layer, is utilized to capture global and local feature information across multiple levels. Experimental results indicate that an accuracy rate of 99.6% is achieved by the HMFF model on the NEU-DET dataset, and 96.9% on the GC10-DET dataset, reflecting improvements of 2.1, 3.8, 3.1, and 2.2% over the ConvNeXt, Swin Transformer, VGG16, and ResNet34 models, respectively, and 4.8, 6.0, 2.6, and 2.8% over other advanced models. The HMFF model outperforms the other models in terms of performance.