Exploiting DCGAN Generated Images for Improving Image Classification

Abstract

The power of Deep Neural Networks relies heavily on the quantity and quality of training data. Usually, it is expensive and time-consuming for people to collect and annotate data on a large scale. Traditional methods, including data augmentation, do not always have the effect, especially in some biomedical fields where some large-size anonymous datasets are generally not publicly available. The study investigates the effectiveness of Deep Convolutional Generative Adversarial Networks (DCGAN)-generated pseudo data compared to simple data augmentation techniques, such as geometric transformations and color enhancements. Various classifiers are trained and tested on both original and augmented datasets. Results indicate that while DCGAN-generated data visually resembles real images, it may not fully capture the statistical characteristics of the original data, leading to decreased classification accuracy compared to simple data augmentation, but the accuracy difference with the original dataset is less than 5% at worst. This shows that the fake data generated by DCGAN can indeed be used for training, but different datasets require network deepening, hyperparameter adjustment, etc.