基于多损失融合与谱归一化的图像超分辨率方法

Multi-loss ensemble and spectral normalization for image super-resolution

图像超分辨率重建研究存在结果客观衡量指标不断变优,但是视觉感知质量依旧平滑的问题。同时,应用生成对抗网络的超分辨率模型中的鉴别器(discriminator)设计存在一个普遍的问题,即训练不稳定问题。针对以上问题作出两点改进:提出多损失融合的方法,寻求一种在PSNR指标与感知质量之间的平衡,通过将均方误差损失、感知损失、风格损失与对抗损失进行融合的方法,在提高PSNR值的同时,改善图像视觉质量;在基于生成对抗网络的超分辨率模型的鉴别器设计中引入谱归一化(spectral normalization),以实现更稳定有效的训练。结果显示,改进后的方法得到了更高的PSNR指标与更逼真的视觉感知质量,并进一步表明感知质量对于超分辨率重建的重要性。

Recently,the objective measurement index of image super-resolution has been improved continuously,but the quality of visual perception is still smooth.And there is a general problem with the discriminator design in the application of the superresolution model,which is the instability of its training.Two improvements are made to the above problems.One was proposed a method of multi-loss ensemble to seek a balance between PSNR indicators and perceived quality.By blending the mean square error loss,perceptual loss,style loss and adversarial loss,it improved the PSNR value while improved the visual quality.The second was to apply spectral normalization in the discriminator design of the GAN-based super-resolution model to achieve more stable and effective training.The results show that the improved method yields a higher PSNR indicator and a more realistic visual perception quality,and further demonstrates the importance of perceived quality for super-resolution reconstruction.