基于改进超分辨率生成对抗网络的图像重建算法

Image Reconstruction Algorithm Based on Improved Super-Resolution Generative Adversarial Network

针对现有基于像素损失的超分辨率图像重建算法对纹理等高频细节的重建效果差问题,提出了一种基于改进超分辨率生成对抗网络(SRGAN)的图像重建算法。首先,去除了生成器中的批归一化层,并结合多级残差网络和密集连接,用残差套残差密集块提高了网络提取特征的能力。然后,结合均方误差与感知损失作为指导生成器训练的损失函数,既保留了图像的高频细节,又避免了伪影的出现。最后,去除了判别器的最后一层Sigmoid层,以更好地收敛训练过程,并用相对损失函数指导判别器的训练。在COCO数据集上的实验结果表明,相比原始SRGAN算法,本算法在Set5数据集上的峰值信噪比(PSNR)、结构相似性(SSIM)分别提高了0.86dB、0.0123;在Set14数据集上的在PSNR、SSIM分别提高了0.69dB、0.0090,且本算法的平均意见指数和视觉效果远优于其他算法。

Aiming at the problem that the existing pixel loss-based super-resolution image reconstruction algorithms have poor reconstruction effect on high-frequency details,such as textures,an image reconstruction algorithm based on an improved super-resolution generative adversarial network(SRGAN)is proposed in this paper.First,remove the batch normalization layers in the generator,combine the multi-level residual network and dense connections,and use the residual-in-residual dense blocks to improve the network’s ability for feature extraction.Then,the mean square error and perceptual loss are combined as the loss function to guide the generator training,which preserves the image’s high-frequency details and avoids the artifacts’appearance.Finally,the last Sigmoid layer of the discriminator is removed to better converge the training process,and the relativistic loss function is used to guide the discriminator training.The experimental results on the COCO dataset show that compared with the original SRGAN algorithm,the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of the algorithm in the Set5 data set are increased by 0.86 dB and 0.0123,respectively,in the Set14 data set,the PSNR and SSIM of the algorithm are improved by 0.69 dB and 0.0090,respectively.The mean opinion index and visual effect of the algorithm are far better than other algorithms.