Using Generative Adversarial Networks Based on Dual Attention Mechanism to Generate Face Images
Yang Yu, Lei Sun, Xiuqing Mao, Leyu Dai, Song Guo, Peiyuan Liu · 2021
Traditional Generative Adversarial Networks (GAN) simulates the correlation between different regions through multiple convolution, but correlation between long-distance is relatively small, resulting in unclear details of generated images. To solve this problem, we propose a model based on dual attention mechanism which called DA-GAN. The self-attention mechanism can extract the dependency effectively between the local and global features of images, allowing the network to learn the connections purposefully between the features, thereby generating higher quality images. The channel-attention mechanism obtains the importance degree of each feature's channel automatically, and improves useful features and suppresses the features that are not useful for the current task according to the importance degree, so as to calculate resources more efficiently. Experiments demonstrate that our method achieves better performance than other method on the CelebA dataset and can produce higher-quality images.