Scholarlink

Deep learning’s convolutional neural networks (CNNs) require extensive, high-quality datasets for optimal performance. When working with limited data samples, these networks often struggle with overfitting and poor generalization. While traditional data augmentation methods help address this challenge by maximizing existing training samples, they offer restricted variation possibilities. This research explores an alternative approach using generative adversarial networks (GANs) for synthetic data creation, specifically focusing on the understudied area of GAN-based data augmentation techniques. Our methodology implements a deep convolutional GAN (DC GAN) architecture for generating synthetic cat facial images. The experimental results demonstrate the efficacy of this approach in two key aspects: enhancing classification accuracy and achieving comparable performance metrics. Through comparative analysis, we found that CNN models trained on a combination of original and GAN-augmented data performed similarly to those trained on substantially larger conventional datasets.

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

Data Augmentation Using Deep Convolutional Generative Adversarial Network for Improving CNN Generalization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Soumya Kukreti,
  • Siddharth,
  • Anuranjana,
  • Sanmukh Kaur

摘要

Deep learning’s convolutional neural networks (CNNs) require extensive, high-quality datasets for optimal performance. When working with limited data samples, these networks often struggle with overfitting and poor generalization. While traditional data augmentation methods help address this challenge by maximizing existing training samples, they offer restricted variation possibilities. This research explores an alternative approach using generative adversarial networks (GANs) for synthetic data creation, specifically focusing on the understudied area of GAN-based data augmentation techniques. Our methodology implements a deep convolutional GAN (DC GAN) architecture for generating synthetic cat facial images. The experimental results demonstrate the efficacy of this approach in two key aspects: enhancing classification accuracy and achieving comparable performance metrics. Through comparative analysis, we found that CNN models trained on a combination of original and GAN-augmented data performed similarly to those trained on substantially larger conventional datasets.