针对图像盲超分辨率网络计算参数多、模型庞大的问题,对快速且节省内存的轻量级图像非盲超分辨率网络(fast and memory-efficient image super resulotion network,FMEN)进行改进,提出了一种轻量级的快速且节省内存的图像盲超分辨率网络(fast and memory-efficient image blind super resulotion network,FMEBN).首先,通过图像退化模块模拟部分真实世界退化空间,使用退化预测模块预测低分辨率(low resolution,LR)图像的退化参数;然后,为能有效利用退化先验信息指导并约束网络进行重建,使用动态卷积对原网络特征提取、重建模块、高频注意力块(high frequency attention block,HFAB)结构进行改进;最后,使用生成对抗网络(generative adversarial network,GAN)对FMEN训练策略与损失函数进行优化,减小真实数据与生成数据的差异,生成更加真实、清晰的纹理、轮廓.实验结果表明,在合成图像数据集和真实图像数据集RealWorld-38上,该算法有较好的重建精度与视觉效果,模型大小 12 MB,可以满足图像盲超分辨率网络的轻量级需求.
A lightweight blind super resulotion network based on generative adversarial network
In order to solve the problem that the image blind super resolution network has many computing parameters and a large model,this paper improves the lightweight image blind super resolution network fast and memory-efficient image super resulotion network(FMEN)and proposes a lightweight image blind super resolution network fast and memory-efficient image blind super resulotion network(FMEBN).Firstly,part of real world degradation space is simulated by image degradation module,and degradation prediction module is used to predict low resolution(LR)image degradation parameters.Then,in order to effectively use degraded prior information to guide and constrain the network reconstruction,dynamic convolution is used to improve the feature extraction,reconstruction module and high frequency attention block(HFAB)structure of the original network.Finally,generative adversarial network(GAN)is used to optimize the training strategy and loss function of FMEN,reduce the difference between real data and generated data,and generate more real and clear texture and contour.The experimental results show that the proposed algorithm has better reconstruction accuracy and visual effect on the synthesized image data set and the real image data set RealWorld-38.The model size is 12 MB,which can meet the lightweight demand of image blind super resolution network.results show that the proposed algorithm has better reconstruction accuracy and visual effect on the synthesized image data set and the real image data set RealWorld-38.The model size is 12 MB,which can meet the lightweight demand of image blind super resolution network.
国家科技期刊平台
NSTL
万方数据
维普
