Generative Adversarial Network for Building Large Custom Dataset

Acquiring new and clean data is a crucial part of building a good machine learning model, especially in industrial fields, where the dataset is usually small and highly imbalance. In this study, we leverage Generative adversarial network (GAN) to extend and enhance our specific dataset by automatically generating and labeling new images. Our pipeline guarantees the success of the synthesis data by turning and stablizing the model’s parameters. Experimental results prove the extraordinary of our dataset in terms of quantitative and qualitative assessment.