Exploiting Deep Convolutional Generative Adversarial Network Generated Images for Enhanced Image Classification

The power of deep neural networks relies heavily on the quantity and quality of training data. However, it is expensive and time consuming for people to collect and annotate data on a large scale. Traditional methods, including modifying the copies of existing data, do not always have the effect, especially in some biomedical fields where some large-size anonymous datasets are generally not publicly available. So, this paper tried to tackle this problem by generating specific data using Deep Convolutional Generative Adversarial Network (DCGAN). DCGAN structure combines convolution and traditional generative adversarial network, has the advantages of producing the clearer images than vanilla Generative adversarial network (GAN). The training dataset is from CIFAR-10 dataset, consist of 10 classes of natural item images. To measure whether it is useful, three classifiers, LeNet, AlexNet and InceptionNet, are trained by feeding original dataset and original dataset mixed with generated data. The final result is presented by comparing accuracy. It goes well by adding more generated data from DCGAN into the original data. The result proves that DCGAN is able to augment data.