Image Classification Method Based on Generative Adversarial Network

摘要

Image classification algorithms based on deep learning often require a large amount of training data with labels, but some times it is difficult to obtain so much training data that meet the requirements. The current methods for data augmentation only operate on the original image and does not change the deep information of the image, so the improvement of models effectiveness is limited. Referring to the idea of CGAN (Conditional Generative Adversarial Networks), we propose an image classification method based on WGAN (Wasserstein GAN). We adding category labels to the data, then sent it into WGAN's generator to training them. Finally, the generator can output the specified category samples, and the ability of discriminator is optimized. Since the discriminator has a part to extract image features, Softmax classifier is added to the last layer of the discriminator so that it can output the category of image and whether it is true or false. Both real samples and generated samples are sent to the discriminator during training, so the number of training samples is increased, the accuracy and robustness of classification model are improved, and the convergence speed of network are accelerated. Experiments on MNIST and CIFAR-10 data sets demonstrate the effectiveness of our method.