Data Augmentation Based on Generative Adversarial Network with Mixed Attention Mechanism

Some downstream tasks often require enough data for training in deep learning, but it is formidable to acquire data in some particular fields. Generative Adversarial Network has been extensively used in data augmentation. However, it still has problems of unstable training and low quality of generated images. This paper proposed Data Augmentation Based on Generative Adversarial Network with Mixed Attention Mechanism (MA-GAN) to solve those problems. This method can generate consistent objects or scenes by correlating the remote features in the image, thus improving the ability to create details. Firstly, the channel-attention and the self-attention mechanism are added into the generator and discriminator. Then, the spectral normalization is introduced into the generator and discriminator so that the parameter matrix satisfies the Lipschitz constraint, thus improving the stability of the model training process. By qualitative and quantitative evaluations on small-scale benchmarks (CelebA, MNIST, and CIFAR-10), the experimental results show that the proposed method performs better than other methods. Compared with WGAN-GP (Improved Training of Wasserstein GANs) and SAGAN (Self-Attention Generative Adversarial Networks), the proposed method contributes to higher classification accuracy, indicating that this method can effectively augment the data of small samples.