Scholarlink

<p>Ensuring defect-free castings is essential for maintaining product reliability and reducing manufacturing losses. Manual inspection is labor-intensive, error-prone, and difficult to scale in modern production lines, motivating the adoption of automated, intelligent inspection systems. This paper presents a lightweight HorNet-inspired Convolutional Neural Network (CNN) for real-time casting defect detection, featuring a Spatially Gated Depthwise Residual (SGDR) block that integrates recursive gated attention with depthwise separable convolutions. This design enhances spatial selectivity and defect localization capability while maintaining extremely low parameter counts and computational costs. The proposed approach is evaluated on a publicly available industrial casting dataset containing both defective and non-defective grayscale images, the same dataset used for all comparative methods to ensure a fair and direct evaluation. Experimental results demonstrate an accuracy of 99.23%, precision of 99.24%, recall of 99.23%, and an F1-score of 99.23%, achieving competitive or superior performance compared to state-of-the-art methods while using only 57,628 parameters (0.22&#xa0;MB) and 0.704&#xa0;GFLOPs for a 224<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>224 input. Grad-CAM visualizations confirm the model’s ability to accurately localize defect-prone regions, enhancing interpretability and operator trust. With its compact size, novel SGDR module, and balanced performance, the proposed network is well-suited for deployment in edge devices and resource-constrained industrial environments. Future work will focus on extending the framework to multi-class defect classification and validating its performance in diverse manufacturing settings.</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 low-complexity spatially gated depthwise residual network for automated visual inspection of castings

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Anju Thomas,
  • Aravind V,
  • B. Chandrababu Naik,
  • Varun P. Gopi

摘要

Ensuring defect-free castings is essential for maintaining product reliability and reducing manufacturing losses. Manual inspection is labor-intensive, error-prone, and difficult to scale in modern production lines, motivating the adoption of automated, intelligent inspection systems. This paper presents a lightweight HorNet-inspired Convolutional Neural Network (CNN) for real-time casting defect detection, featuring a Spatially Gated Depthwise Residual (SGDR) block that integrates recursive gated attention with depthwise separable convolutions. This design enhances spatial selectivity and defect localization capability while maintaining extremely low parameter counts and computational costs. The proposed approach is evaluated on a publicly available industrial casting dataset containing both defective and non-defective grayscale images, the same dataset used for all comparative methods to ensure a fair and direct evaluation. Experimental results demonstrate an accuracy of 99.23%, precision of 99.24%, recall of 99.23%, and an F1-score of 99.23%, achieving competitive or superior performance compared to state-of-the-art methods while using only 57,628 parameters (0.22 MB) and 0.704 GFLOPs for a 224 \(\times \) × 224 input. Grad-CAM visualizations confirm the model’s ability to accurately localize defect-prone regions, enhancing interpretability and operator trust. With its compact size, novel SGDR module, and balanced performance, the proposed network is well-suited for deployment in edge devices and resource-constrained industrial environments. Future work will focus on extending the framework to multi-class defect classification and validating its performance in diverse manufacturing settings.