基于反馈生成对抗网络的单图像超分辨率重建

Single image super-resolution reconstruction based on feedback generative adversarial network

针对现有图像超分辨率重建方法中高频图像信息不丰富的问题,提出一种基于反馈和注意机制的单图像重建生成对抗网络(GFSRGAN)。采用反馈网络作为生成器,通过反馈连接逐步生成高分辨率图像;提出一种具有注意机制的反馈块,其能在处理反馈流的同时,自适应地选择有用的特征信息;利用相对平均最小二乘GAN(RaLSGAN)损失引导模型获得更真实的图像。实验结果表明,与现有基于GAN的超分辨方法相比,该方法重建出的图像纹理更加逼真自然。

A single-image reconstruction generative adversarial network(GFSRGAN)based on feedback and attention mechanisms was proposed to address the problem that the existing image super-resolution reconstruction methods are not rich in high-frequency image information.A feedback network was used as a generator,which gradually generated high-resolution images through feedback connections.A feedback block with an attention mechanism was proposed,which was able to process the feedback stream while adaptively selecting useful feature information.A more realistic image was obtained using a relative mean least squares GAN(RaLSGAN)loss-guided model.Experimental results show that the reconstructed image texture is more realistic and natural compared with the existing GAN-based super-resolution methods.