CIGAN: A Novel GANs Model Based on Category Information

Abstract

Unsupervised learning with Generative Adversarial Networks (GANs) has made great achievements these years. Generally, traditional GANs model improve the generation effect by proposing new loss functions or adjusting network structures. However, the category information of training data is rarely used in the training of GANs network and the discriminative power of traditional GANs' discriminator is limited. To overcome such a problem, we propose an improved GANs model based on Category Information (CIGAN), which applies the hash center loss to improve the training speed of the CIGAN model. We also propose two methods for CIGAN model to improve the discrimination ability of discriminator network. The CIGAN model has advantages over traditional GANs model as it can generate higher quality images, and can remain stable during the learning process when the batch normalization layer is removed. The Inception score on the CIFAR-10 dataset also demonstrates that our model is better than traditional GANs model with higher generation effect.