Scholarlink

<p>To address the limited availability of atmospheric natural environment test samples and the challenge of obtaining sufficient industrial damage sample data for deep learning, we propose IDS-GAN, a novel data generation method based on an improved DCGAN for small-sample industrial damage modeling. First, a Wasserstein distance loss function with a gradient penalty term is introduced as an adversarial loss to mitigate mode collapse and stabilize training. In addition, L1 loss and Structural Similarity Index (SSIM) loss are incorporated to guide the generator in refining local details in high-resolution images. Residual blocks are added to the generator to address the gradient vanishing problem commonly encountered in deep networks during high-resolution image generation. Experimental results demonstrate that, compared to a progressively trained DCGAN, the proposed IDS-GAN model generates industrial damage sample images with significantly higher quality and more distinct features. Specifically, the Fréchet Inception Distance (FID) score decreased by 18.6%, while the Peak Signal-to-Noise Ratio (PSNR) and Inception Score (IS) improved by 5.1% and 21.1%, respectively. These results indicate that IDS-GAN is an effective approach for generating industrial damage sample datasets, offering potential applications in relevant research fields.</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

Industrial Damage Sample Image Generation via IDS-GAN with Multi-Loss Optimization and Residual Learning

  • 期刊论文

  • 发表日期：

    2025年7月4日

查看全文
  • Hongfei Li,
  • Xudong Wang,
  • Wengang Ao

摘要

To address the limited availability of atmospheric natural environment test samples and the challenge of obtaining sufficient industrial damage sample data for deep learning, we propose IDS-GAN, a novel data generation method based on an improved DCGAN for small-sample industrial damage modeling. First, a Wasserstein distance loss function with a gradient penalty term is introduced as an adversarial loss to mitigate mode collapse and stabilize training. In addition, L1 loss and Structural Similarity Index (SSIM) loss are incorporated to guide the generator in refining local details in high-resolution images. Residual blocks are added to the generator to address the gradient vanishing problem commonly encountered in deep networks during high-resolution image generation. Experimental results demonstrate that, compared to a progressively trained DCGAN, the proposed IDS-GAN model generates industrial damage sample images with significantly higher quality and more distinct features. Specifically, the Fréchet Inception Distance (FID) score decreased by 18.6%, while the Peak Signal-to-Noise Ratio (PSNR) and Inception Score (IS) improved by 5.1% and 21.1%, respectively. These results indicate that IDS-GAN is an effective approach for generating industrial damage sample datasets, offering potential applications in relevant research fields.