摘要
超分辨率作为一种经典的视觉任务,在多个领域有着广泛的应用。随着深度学习中无监督学习的发展,以及生成对抗网络(generative adversarial network,GAN)的提出,超分辨率技术又得到了进一步的提高,但是相关网络仍旧存在过拟合、泛化性弱等诸多问题。以超分辨率生成对抗网络(super-resolution generative adversarial network,SRGAN)为基础,受研究Dropout在经典超分辨率网络中影响的相关论文启发,在SRGAN中加入Dropout层并研究其对生成图像质量的影响,采用峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性(structural similarity,SSIM)评估图像质量。实验结果表明在合适的Dropout参数下,网络重建图像具有更好的视觉效果,PSNR能够达到0.4左右的提升同时SSIM也有提高,从训练过程中不同迭代次数生成图像的比较发现改进后网络缓解了训练不稳定问题。将Dropout层加入超分辨率网络不同于以往的方法,为改进此类网络提供了一个新的思路。
As a classic visual task,super-resolution has a wide range of applications in many fields.With the development of unsu-pervised learning in deep learning and the proposal of GANs in generative adversarial networks,super-resolution technology has been further improved,but related networks still have many problems such as overfitting and weak generalization.Based on SRGAN,in-spired by the relevant papers studying the influence of Dropout in classical super-resolution networks,layer was added to SRGAN and its influence on the quality of the generated image was studied.The image quality was evaluated with the peak signal-to-noise ratio PSNR(peak signal to noise ratio)and structural similarity SSIM(structural similarity).And the experimental results show that the network reconstructed image has better visual effects under suitable dropout parameters.The PSNR value can reach an increase of about 0.4,and SSIM is also improved,and from the comparison of images generated by different iterations in the training process,it is found that the improved network alleviates the problem of training instability.Adding the layer to a super-resolution network is different from previous methods and provides a new way to improve such networks.
作者
刘慧
卢云志
张雷
LIU Hui;LU Yun-zhi;ZHANG Lei(Electrical&Information Engineering School,Beijing University of Civil Engineering and Architecture,Beijing 102616,China)
出处
《科学技术与工程》
北大核心
2023年第23期10015-10022,共8页
Science Technology and Engineering
