Abstract:

Despite the deep convolutional neural networks make breakthroughs in accuracy and speed of single image super-resolution, the central problems remain largely unsolved: the reconstruction results for traditional methods which are not obvious and too smooth. In this paper, a single image super resolution method based on coupled generative adversarial networks is proposed to solve the central problems. The deep residual network and the deep convolutional neural network are used as the generator and the discriminator. The attention augmented convolution is applied to the generator. In order to enhance the quality of the generated image and the stability of the training process, strike the right balance between the capabilities of generators and discriminators. The relative discriminator is used to calculate the loss value from adversarial neural network. Compared with the mainstream super-resolution reconstruction algorithm on the basic of Set5, Set14 and BSD100 datasets, the experimental results show that the proposed approach achieves consistently better visual quality with more sharper, realistic and high-frequency textures than prior methods, and is effective in diversified image generation applications.

Key words: super-resolution, generative adversarial network, attention augmented, deep residual network, diversity

摘要：

尽管卷积神经网络在实现单帧图像超分辨率的准确性和速度方面取得一定突破，但仍然存在重建结果细节不明显，过于光滑等中心问题。针对这一中心问题，提出一种基于单帧图像的耦合生成式对抗超分辨率重建算法，定义的生成器和判别器分别采用深度残差网络和深度卷积网络，将自注意力增强卷积应用到生成器网络中，为了增强生成图像的质量和训练过程的稳定，对生成器和判别器的学习能力进行平衡，使用相对判别器计算来自对抗神经网络的损失值。主流超分辨重建算法在Set5、Set4、BSD100经典数据集上进行对比，实验结果表明，提出的算法在边缘锐化、真实性和获得更好的高频细节恢复方面能够达到更好的连续视觉效果，同时能够增强生成图像的多样性。

关键词: 超分辨率, 生成对抗网络, 自注意力增强, 深度残差网络, 多样性