Scholarlink

<p>Artificial intelligence-based defect detection relies heavily on accurate and efficient algorithms to operate in real time. However, a significant challenge lies in obtaining sufficient and diverse datasets for training and testing these algorithms, particularly when utilizing deep learning (DL) techniques. Accessing large-scale datasets, especially in manufacturing environments, presents obstacles such as expensive data collection hardware and the time needed to collect all the required data. This paper introduces a novel solution, leveraging Blender with its Python Application Programming Interface (API), to efficiently generate diverse and realistic datasets for training three different convolutional neural network (CNN)-based models (Custom CNN, VGG16, and EfficientNet) in defect detection applications in water bottles. The results showed that the EfficientNet model scored the highest performance with an accuracy of 99.94% but at the expense of computational cost, while the Custom CNN model performed less accurately (98.87% accuracy) while maintaining a suitable edge device performance. This approach circumvents the limitations associated with traditional data collection methods by synthesizing images of defective and non-defective bottles. The generated datasets enable the training of CNN-based models capable of accurately detecting defects, optimizing quality control processes, and minimizing waste throughout the manufacturing lifecycle. This innovative methodology promises to revolutionize defect detection by providing a cost-effective, scalable, and sustainable solution to address the challenges of dataset acquisition and algorithm training.</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

Toward sustainable production: a synthetic dataset framework to accelerate quality control via generative and predictive AI

  • 期刊论文

  • 发表日期：

    2025年6月11日

查看全文
  • Mohammad Shahin,
  • F. Frank Chen,
  • Mazdak Maghanaki,
  • Hamed Mehrzadi,
  • Ali Hosseinzadeh

摘要

Artificial intelligence-based defect detection relies heavily on accurate and efficient algorithms to operate in real time. However, a significant challenge lies in obtaining sufficient and diverse datasets for training and testing these algorithms, particularly when utilizing deep learning (DL) techniques. Accessing large-scale datasets, especially in manufacturing environments, presents obstacles such as expensive data collection hardware and the time needed to collect all the required data. This paper introduces a novel solution, leveraging Blender with its Python Application Programming Interface (API), to efficiently generate diverse and realistic datasets for training three different convolutional neural network (CNN)-based models (Custom CNN, VGG16, and EfficientNet) in defect detection applications in water bottles. The results showed that the EfficientNet model scored the highest performance with an accuracy of 99.94% but at the expense of computational cost, while the Custom CNN model performed less accurately (98.87% accuracy) while maintaining a suitable edge device performance. This approach circumvents the limitations associated with traditional data collection methods by synthesizing images of defective and non-defective bottles. The generated datasets enable the training of CNN-based models capable of accurately detecting defects, optimizing quality control processes, and minimizing waste throughout the manufacturing lifecycle. This innovative methodology promises to revolutionize defect detection by providing a cost-effective, scalable, and sustainable solution to address the challenges of dataset acquisition and algorithm training.