Scholarlink

<p>The proposed research provides a structured approach for automation of 3D printing quality inspection through the comparative analysis of three deep learning models: single shot detector (SSD), Detectron2, and you only look once (YOLO) version 11. Using 1000 images of 3D-printed keychains–650 defective and 350 intact–the study evaluates detection accuracy, preprocessing needs, and computational efficiency of each architecture. The evaluation reveals distinct performance characteristics: YOLO v11 achieves optimal accuracy (94 %) with flexibility in both preprocessed and raw image analysis, making it applicable for real-time quality control operations. Detectron2 exhibits exceptional versatility with 90 % accuracy in both preprocessing states, providing a stable solution for manufacturing systems with a high variability of conditions. SSD demonstrates effective results (89 % accuracy) but exhibits reliance on preprocessing techniques, notably edge detection in the Canny sense, to maintain performance levels. The implementation incorporates controlled image capture protocols, standardized preprocessing methods, and extensive training procedures optimized for keychain-specific features. Results demonstrate the practical applicability of deep learning integration in additive manufacturing quality assurance, with model selection requirements tied to manufacturing constraints regarding computational resources, required detection speed, and operational flexibility. The analysis provides a scientific basis for model selection in production systems that require constant high-throughput inspection with a reasonable degree of flexibility and accuracy.</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

Automated quality inspection of 3D printed components using deep learning models: A comparative analysis of YOLO, SSD, and detectron

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Sanjana Nithyananda Kumar,
  • Immanual Rajkumar,
  • David Rathnaraj J.

摘要

The proposed research provides a structured approach for automation of 3D printing quality inspection through the comparative analysis of three deep learning models: single shot detector (SSD), Detectron2, and you only look once (YOLO) version 11. Using 1000 images of 3D-printed keychains–650 defective and 350 intact–the study evaluates detection accuracy, preprocessing needs, and computational efficiency of each architecture. The evaluation reveals distinct performance characteristics: YOLO v11 achieves optimal accuracy (94 %) with flexibility in both preprocessed and raw image analysis, making it applicable for real-time quality control operations. Detectron2 exhibits exceptional versatility with 90 % accuracy in both preprocessing states, providing a stable solution for manufacturing systems with a high variability of conditions. SSD demonstrates effective results (89 % accuracy) but exhibits reliance on preprocessing techniques, notably edge detection in the Canny sense, to maintain performance levels. The implementation incorporates controlled image capture protocols, standardized preprocessing methods, and extensive training procedures optimized for keychain-specific features. Results demonstrate the practical applicability of deep learning integration in additive manufacturing quality assurance, with model selection requirements tied to manufacturing constraints regarding computational resources, required detection speed, and operational flexibility. The analysis provides a scientific basis for model selection in production systems that require constant high-throughput inspection with a reasonable degree of flexibility and accuracy.