基于生成对抗网络的图像超分辨率重建算法

Image super resolution reconstruction algorithm based on generative countermeasure network

SRGAN是基于深度学习的图像超分辨率的典型方法,重建效果较好,但该算法还存在一些缺陷,在提高图像质量和运行速度上仍然有较大提升空间。本文在SRGAN网络模型的基础上提出了一个优化模型。因为批量归一化(BN)层在超分辨图像重建中常常会忽略一些图像的细节,同时增加网络的复杂度,所以在SRGAN的生成器中去除了BN层,并引入ECA通道注意力,使每个残差块生成特征图获得相应的权重,以便处理更多的图像细节。经过公开数据集的训练和对比实验,结果表明提出的改进模型相比于对比模型,重建图像的细节恢复更丰富,视觉效果更好,峰值信噪比和结构相似性表现更佳,模型总参数量更少。

SRGAN is a typical method of image super-resolution based on deep learning,the reconstruction effect is good,but the algorithm still has some shortcomings,and there is still more room for improving the image quality and operation speed.An optimization model is proposed based on the SRGAN network model.Because the batch normalization(BN)layer often ignores some image details in super-resolution image reconstruction and increases the complexity of the network at the same time,the BN layer is removed from the generator of SRGAN and the ECA channel attention is introduced so that each residual block generating feature map gets a corresponding weight in order to process more image details.After training and comparison experiments on public datasets,the results show that the proposed improved model has richer image details recovery,better visual effects,better peak signal-to-noise ratio and structural similarity performance,and fewer total number of model parameters compared to the comparison model.