Scholarlink

<p>To address the inefficiency and instability of manual classification for surface defects in automotive flange disks, this paper proposes an integrated SVM classification framework combining improved Canny edge detection with enhanced Hu moment features. The dataset construction process involves capturing 50 raw images of each category, including sand holes, scratches, and defect-free samples, using an industrial camera. A 4× data augmentation strategy was applied, including mirror flipping, random panning within ± 10 pixels, and gamma correction with span coefficients of 0.5 to 1.5 to generate a standardized dataset of 600 images. These were partitioned into 420 training samples and 180 independent testing samples. During preprocessing, the improved bilateral filtering (gradient similarity instead of grayscale similarity) improves PSNR by 12.4% compared to traditional Gaussian filtering. Extended Sobel operators incorporating 45° and 135° gradient templates improved annular workpiece feature representation, integrated with Otsu-based adaptive dual-threshold segmentation. For feature extraction, a six-dimensional modified Hu moment descriptor was developed through normalized second-order central moments, constructing scale-invariant ρ1-ρ6 feature vectors that eliminate discrete-scale sensitivity. Experimental validation via five-fold cross-validation using 480 training samples per fold optimized SVM parameters, implementing a third-order polynomial kernel with a penalty factor C = 76.8. The model demonstrated 97.2%±0.6% classification accuracy on independent test sets, surpassing baseline SVM by 13.5% while maintaining efficient inference speed of 0.684&#xa0;s per image. Comparative analysis shows that the performance is better than MobileNetV3, with an increase in accuracy and recall of 0.7% and 1.8%, respectively. Furthermore, its classification accuracy is comparable to that of EfficientNetB0 (97.4%), while its inference speed is significantly faster—approximately 9.1 times. This optimized framework provides methodological advancements and practical insights for industrial quality inspection applications.</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

Research on surface defect classification model of automotive flange disks based on machine vision

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • Jun Guo,
  • Chengyang Ge,
  • Muhammad Sohail Memon,
  • Huimin Fang,
  • Qingyi Zhang,
  • Shengqiang Lin,
  • Liusong Yang

摘要

To address the inefficiency and instability of manual classification for surface defects in automotive flange disks, this paper proposes an integrated SVM classification framework combining improved Canny edge detection with enhanced Hu moment features. The dataset construction process involves capturing 50 raw images of each category, including sand holes, scratches, and defect-free samples, using an industrial camera. A 4× data augmentation strategy was applied, including mirror flipping, random panning within ± 10 pixels, and gamma correction with span coefficients of 0.5 to 1.5 to generate a standardized dataset of 600 images. These were partitioned into 420 training samples and 180 independent testing samples. During preprocessing, the improved bilateral filtering (gradient similarity instead of grayscale similarity) improves PSNR by 12.4% compared to traditional Gaussian filtering. Extended Sobel operators incorporating 45° and 135° gradient templates improved annular workpiece feature representation, integrated with Otsu-based adaptive dual-threshold segmentation. For feature extraction, a six-dimensional modified Hu moment descriptor was developed through normalized second-order central moments, constructing scale-invariant ρ1-ρ6 feature vectors that eliminate discrete-scale sensitivity. Experimental validation via five-fold cross-validation using 480 training samples per fold optimized SVM parameters, implementing a third-order polynomial kernel with a penalty factor C = 76.8. The model demonstrated 97.2%±0.6% classification accuracy on independent test sets, surpassing baseline SVM by 13.5% while maintaining efficient inference speed of 0.684 s per image. Comparative analysis shows that the performance is better than MobileNetV3, with an increase in accuracy and recall of 0.7% and 1.8%, respectively. Furthermore, its classification accuracy is comparable to that of EfficientNetB0 (97.4%), while its inference speed is significantly faster—approximately 9.1 times. This optimized framework provides methodological advancements and practical insights for industrial quality inspection applications.