Scholarlink

<p>This study presents a comparative evaluation of three deep learning models: MobileNetV2, EfficientNetB0, and ResNet152-for defect classification in water transfer printing manufacturing. We assess their performance across six image resolutions (64 × 64 to 600 × 600 pixels), analyzing key metrics such as accuracy, inference time, and computational efficiency. The models were tested on five defect categories: entrapped air, film creasing, film delamination, pattern distortion, and uniform transfer. MobileNetV2 demonstrated the fastest inference times, ranging from 9.6 ms at 64 × 64 resolution to 572 ms at 600 × 600, with test accuracy peaking at 77.13 % (224 × 224) before declining at higher resolutions. EfficientNetB0 offered a strong balance, achieving up to 87.89 % test accuracy at higher resolutions, with inference times increasing from 14 ms (64 × 64) to 823 ms (600 × 600). ResNet152 delivered the highest test accuracy (up to 95.94 % at 600 × 600), but incurred the highest computational cost, with inference times rising from 55.91 ms (64 × 64) to nearly 3 seconds (2979.75 ms) at maximum resolution. These results confirm a non-linear relationship between image resolution and model performance: as resolution increases from 64 × 64 to 600 × 600 pixels, test accuracy improves by 22–53 % depending on the model, while inference times escalate dramatically-by over 5100 % to nearly 5900 % for all architectures.</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

Balancing accuracy and efficiency: comparative analysis of deep learning models for defect classification in water transfer printing

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Gyuseok Lee,
  • Joel Ndikumana,
  • Hyunjoon Choi,
  • Seonghyun Choi,
  • Jaeseok Yang,
  • Kunsik An

摘要

This study presents a comparative evaluation of three deep learning models: MobileNetV2, EfficientNetB0, and ResNet152-for defect classification in water transfer printing manufacturing. We assess their performance across six image resolutions (64 × 64 to 600 × 600 pixels), analyzing key metrics such as accuracy, inference time, and computational efficiency. The models were tested on five defect categories: entrapped air, film creasing, film delamination, pattern distortion, and uniform transfer. MobileNetV2 demonstrated the fastest inference times, ranging from 9.6 ms at 64 × 64 resolution to 572 ms at 600 × 600, with test accuracy peaking at 77.13 % (224 × 224) before declining at higher resolutions. EfficientNetB0 offered a strong balance, achieving up to 87.89 % test accuracy at higher resolutions, with inference times increasing from 14 ms (64 × 64) to 823 ms (600 × 600). ResNet152 delivered the highest test accuracy (up to 95.94 % at 600 × 600), but incurred the highest computational cost, with inference times rising from 55.91 ms (64 × 64) to nearly 3 seconds (2979.75 ms) at maximum resolution. These results confirm a non-linear relationship between image resolution and model performance: as resolution increases from 64 × 64 to 600 × 600 pixels, test accuracy improves by 22–53 % depending on the model, while inference times escalate dramatically-by over 5100 % to nearly 5900 % for all architectures.