Use of Improved Generative Adversarial Network (GAN) Under Insufficient Data

The article covers enhancements to the generative adversarial network (GAN) model’s architecture and training, enabling stable training in the absence of sufficient data. An improved generative adversarial network (GAN) architecture has been proposed. These improvements are then applied to the augmentation of a dataset on tyre joint defects, which is utilised for classification applications. The dataset used has a higher percentage of conformity images and is quite uneven. It is difficult to create precise defect classification models given this uneven and constrained dataset of defect identification. So, in the work that is being presented, research is done to expand the defect dataset and improve the balance between the various defect classifications. Indeed, the quality of generated images has considerably improved as a result of recent developments in generative adversarial networks (GANs). Deep learning models in the GAN class combine a generator network with a discriminator network. The current study reveals that the recommended augmented GAN model is useful in enhancing the performance classification model under a small dataset. The generated effects of progressed GAN are evaluated using the Fréchet Inception Distance (FID) score, which indicates extensive development over the styleGAN architecture. Additional dataset augmentation exams making use of generated photos monitor a 10% boom in category version precision in comparison to the preliminary dataset. To evaluate the effectiveness of GAN-generated picture augmentation, PCA plots can be used to visualize the distribution of real and augmented images in a lower-dimensional space.