MDA-GAN: Multi-dimensional Attention Guided Concurrent-Single-Image-GAN

Recent years have seen great achievements made in learning generative models from a single image, which is groundbreaking for image generation methods. Although some different strategies have been proposed to accomplish single image generation tasks, such as SinGAN and ConSinGAN, two main challenges do exist which are the scarcity of sufficient training samples and the hardship of obtaining all feature information from a single image. In order to train a model that can perfectly extract image features with very few samples, this paper introduces MDA-GAN, the multi-dimensional attention guided GAN for single image generation tasks, which aims to generate high-quality images. Specifically, in the multi-dimensional attention (MDA) module, 1D channel-wise and 2D spatial-wise branches focus on the important contents and specific positions of a single image. Besides, the 3D neuronal-wise branch uses an enhanced energy function to distinguish the differences between neurons and extract their features. We concatenate all feature maps together to facilitate the model focusing on different parts of the whole image and generating realistic details. Finally, our attention module is embedded into a single-image GAN model. We conduct experiments on three benchmark datasets (the Places, LSUN, and mini-ImageNet) to validate its potency. Our approach surpasses state-of-the-art baselines, including SinGAN, ConSinGAN, and CCASinGAN, in generating superior images. Through qualitative and quantitative experiments, MDA-GAN outperforms existing generative models. Our method significantly enhances single-image generative models by addressing dimension-related challenges, crucially contributing to high-quality image generation.