【作者】 王令昌； 李蓉； 张海斌；

【Author】 LingChang Wang;Rong Li;Haibin Zhang;Beijing university of technology;

【机构】 北京工业大学；

【摘要】 本文结合注意力机制与生成对抗网络,设计了一种基于注意力机制的图像超分辨率算法。具体的,本文基于ESRGAN的研究,在原来ESRGAN架构中添加注意力模块。注意力模块识别高频细节位置,通过增强高频细节特征获得了更好的图像重建结果;另外使用残差密集块作为生成器网络的基本结构单元,有效避免了算法过拟合问题。在经典数据集的实验结果证明,所提出的算法能够较好恢复图像的高频细节,提升图像的输出质量。更多还原

【Abstract】 In this paper we present an image super-resolution algorithm which combined with attention mechanism and generative adversarial network(GAN). Specifically, based on the research of ESRGAN, we add an attention module to the framework to identify high-frequency contents. By enhancing high frequency details, better super-resolution results are achieved; The residualin-residual dense block(RRDB) is used as the basic component of the generator network, which effectively avoids the over-fitting problem. The experimental results on the benchmark datasets show that the proposed method have a competitive effect on restoring the high-frequency details of the image.更多还原