GAN-Based Data Augmentation Strategies for Imbalanced Industrial Datasets: A Comparative Study with Transfer Learning

Abstract

Automated surface defect detection in industrial manufacturing faces a critical challenge: the scarcity of balanced, annotated datasets severely limits the performance of deep learning models, particularly for minority defect classes that occur infrequently in production environments. While Generative Adversarial Networks (GANs) have emerged as promising tools for synthetic data generation, systematic comparative studies evaluating different GAN architectures and augmentation strategies for industrial defect detection remain limited. Furthermore, the integration of GAN-generated data with transfer learning techniques has not been thoroughly investigated. This study presents a comprehensive comparative analysis of four prominent GAN architectures, Deep Convolutional GAN (DCGAN), Conditional GAN (CGAN), Auxiliary Classifier GAN (ACGAN), and Wasserstein GAN with Gradient Penalty (WGAN-GP), applied to surface defect classification on a merged industrial dataset combining NEU-CLS and X-SSD (3,160 images across 13 defect classes). We systematically evaluate four augmentation strategies: uniform generation, selective minority oversampling, balanced augmentation, and distribution-aware generation. The generated synthetic images are integrated with transfer learning using a ResNet-50 backbone pretrained on ImageNet to assess downstream classification performance. Results show that GAN-based augmentation improves classification accuracy, particularly for minority defect types, however results depend strongly on the augmentation strategy. CGAN offered the best balance between accuracy, stability, and practical reliability. This work contributes by providing a benchmark for evaluating GAN-based augmentation methods in industrial inspection and by proposing a hybrid GAN–ResNet50 pipeline, which achieves superior classification accuracy through the integration of generative modeling with transfer learning.