Scholarlink

<p>Following the rapid development of artificial intelligence and industrial networking technologies, real-time automated defect detection applications have started to be widely implemented in product manufacturing lines. These automated defect detection applications are replacing traditional manual inspections by quality control personnel and can significantly enhance product quality and operational efficiency. There are practical challenges associated with the training of artificial intelligence defect detection models, such as insufficient training samples and the rarity of their availability, which lead to reduced learning efficiency and generalisation capability; the application of generalised adversarial network (GAN) technology can therefore enhance the diversity of the training samples through the creation of synthetic data, thus improving the training performance of artificial intelligence defect detection models. However, traditional GAN techniques rely on a single defect sample category, resulting in synthetic data that are overly similar to the original data. This issue limits the generalisation capability of artificial intelligence defect detection models and prevents significant improvements in overall recognition accuracy. This paper proposes the use of cycle-consistent generative adversarial network (CycleGAN) to construct a multi-key-feature reference framework, in which the content loss of the generated images is adjusted to strike a balance between the similarity and diversity of the synthetic data, thereby enhancing the learning efficiency and generalisation capability of artificial intelligence defect detection models. Training of the CycleGAN model incorporates a multi-key-feature reference extractor that performs feature vector transformations on both original defect sample category data and synthetic data. The cosine similarity is used to evaluate the content loss between the synthetic data and various original defect sample category data, thus enabling adaptive adjustments to generate training samples that meet customised content loss requirements. This approach achieves a balance between synthetic data and the content loss of various original defect sample category data, to produce suitable training samples for artificial intelligence defect detection models. The framework proposed in this paper yields an improvement in recognition accuracy of at least 5% for artificial intelligence defect detection compared to models in the literature.</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

Content loss adjustment in image generation via a multiple key features reference structure for cycle generative adversarial networks

  • 期刊论文

  • 发表日期：

    2025年10月20日

查看全文
  • Ming-Fong Tsai,
  • Heng-Chih Liu,
  • Shu-Lin Hsieh

摘要

Following the rapid development of artificial intelligence and industrial networking technologies, real-time automated defect detection applications have started to be widely implemented in product manufacturing lines. These automated defect detection applications are replacing traditional manual inspections by quality control personnel and can significantly enhance product quality and operational efficiency. There are practical challenges associated with the training of artificial intelligence defect detection models, such as insufficient training samples and the rarity of their availability, which lead to reduced learning efficiency and generalisation capability; the application of generalised adversarial network (GAN) technology can therefore enhance the diversity of the training samples through the creation of synthetic data, thus improving the training performance of artificial intelligence defect detection models. However, traditional GAN techniques rely on a single defect sample category, resulting in synthetic data that are overly similar to the original data. This issue limits the generalisation capability of artificial intelligence defect detection models and prevents significant improvements in overall recognition accuracy. This paper proposes the use of cycle-consistent generative adversarial network (CycleGAN) to construct a multi-key-feature reference framework, in which the content loss of the generated images is adjusted to strike a balance between the similarity and diversity of the synthetic data, thereby enhancing the learning efficiency and generalisation capability of artificial intelligence defect detection models. Training of the CycleGAN model incorporates a multi-key-feature reference extractor that performs feature vector transformations on both original defect sample category data and synthetic data. The cosine similarity is used to evaluate the content loss between the synthetic data and various original defect sample category data, thus enabling adaptive adjustments to generate training samples that meet customised content loss requirements. This approach achieves a balance between synthetic data and the content loss of various original defect sample category data, to produce suitable training samples for artificial intelligence defect detection models. The framework proposed in this paper yields an improvement in recognition accuracy of at least 5% for artificial intelligence defect detection compared to models in the literature.