Data Generation Methods of Generative Adversarial Networks Based on Exponential Adaptive Pool.

In the case of small sample training, the extraction of image features is crucial to the quality of generated images in Generative Adversarial Network (GAN). To ensure it, convolution and pooling are often used to increase the depth of the network, while inevitably lead to the loss of features. To address the problem above, this paper proposes an innovative model of GAN named EAP-GAN based on exponential adaptive pool. To further improve the stability of training process and the quality of generated images, exponential adaptive pool is introduced to replace the average pooling originally, which calculates the kernel weights in two ways: the distance from the centre vector and the size of neighbouring pixel values in the kernel region and then adaptively fuses the feature maps obtained in these two ways by means of the learnable parameter beta, thus presevering the image details more effectively. Finally, the EAP-FastGAN model is verified to outperform the original FastGAN model by visualizing loss functions and comparing the values of FID (a classical quantitative method) on four datasets, namely Panda, Anime, Dog and Obama.