基于改进生成对抗网络的单帧图像超分辨率重建

Single frame image super-resolution reconstruction based on improved generative adversarial network

为了获得更好的图像超分辨率重建质量,提高网络训练的稳定性,对生成对抗网络、损失函数进行研究。首先,介绍了SRGAN和DenseNet,并设计了基于DenseNet的生成网络用以生成图像,且将子像素卷积模块加入到DenseNet中。接着,移除了原本DenseNet中冗余的BN层,提高了模型的训练效率。最后,介绍了SRGAN的损失函数并基于Earth-Mover距离来重新设计损失函数,并且用SmoothL1损失取代MSE损失来计算VGG特征图,以防止MSE放大最大误差和最小误差间的差距。实验证明该模型在网络训练过程中能够达到稳定收敛的状态。重建出的图像质量对比SRGAN,在3个基准测试集SET5,SET14,BSD100上的平均PSNR要高约2.02 dB,SSIM高约0.042(5.6%)。重建出的图像不仅在指标上有所提升,且拥有更好的清晰度,高频细节更为丰富。

In order to obtain better image super-resolution reconstruction quality and improve the stability of network training,the generation of confrontation networks and loss functions are studied.Firstly,SRGAN and DenseNet are introduced,a generation network is designed to generate image based on DenseNet,and the sub-pixel convolution module is added to DenseNet.Then,the redundant BN layer in the original DenseNet is removed to improve the training efficiency of the model.Finally,the loss function of SRGAN is introduced and the loss function is redesigned based on the Earth-Mover distance,and the SmoothL1 loss is used to replace the MSE loss to calculate the VGG feature map to prevent MSE from amplifying the gap between the maximum error and the minimum error.Experiments prove that the model can achieve a stable convergence state during the network training process.The quality of the reconstructed image is compared with SRGAN,the average PSNR on the three benchmark test sets SET5,SET14,and BSD100 is about 2.02 dB higher,and SSIM is about 0.042(5.6%)higher.The reconstructed image not only has improved indicators,but also has better definition and richer high-frequency details.