Generative Adversarial Network Optimization Algorithm Based on Adaptive Data Augmentation

In the research of deep learning, an Unsupervised deep convolution-generated Generated Adversarial Network (UGAN) usually needs a large number of data samples to train. However, when faced with some small samples, the performance of the algorithm is often degraded due to over-fitting. Combined with specially designed data enhancement methods, a generated adversarial network optimization algorithm based on adaptive data augmentation (AdauGAN) is proposed. The adaptive data augmentation module is added before the discriminant network, and a spatial transformation is carried out simultaneously at the probability distribution level of generated data and real data. To alleviate the over-fitting phenomenon in the training process, the current enhancement intensity is adjusted adaptively after the over-fitting occurs. The proposed algorithm is verified on SVHN, CelebA and CIFAR-10 data sets. The Frechet Inception Distance (FID) values of AdauGAN achieve 22.10, 23.94, 34.87, respectively, which is close to or even higher than the training results of Deep Convolution Generated Adversarial Network (DCGAN) under all data. Extensive experiment results show that the proposed Adaugan has an excellent performance in small samples. Besides, in some cases, it can catch up with the large sample results of existing algorithms.