Abstract
Although SISR (Single Image Super Resolution) problem can be effectively solved by deep learning based methods, the training phase often considers single degradation type such as bicubic interpolation or Gaussian blur with fixed variance. These priori hypotheses often fail and lead to reconstruction error in real scenario. In this paper, we propose an end-to-end CNN model RPSRMD to handle SR problem in multiple Gaussian degradations by extracting and using as side information a shared image prior that is consistent in different Gaussian degradations. The shared image prior is generated by an AED network RPGen with a rationally designed loss function that contains two parts: consistency loss and validity loss. These losses supervise the training of AED to guarantee that the image priors of one image with different Gaussian blurs to be very similar. Afterwards we carefully designed a SR network, which is termed as PResNet (Prior based Residual Network) in this paper, to efficiently use the image priors and generate high quality and robust SR images when unknown Gaussian blur is presented. When we applied variant Gaussian blurs to the low resolution images, the experiments prove that our proposed RPSRMD, which includes RPGen and PResNet as two core components, is superior to many state-of-the-art SR methods that were designed and trained to handle multi-degradation.
Highlights
As one of the most active research topics, single image superresolution (SISR) has attracted increasingly attention
Since most of the existing SR methods assumed that the LR image is down-sampled by some pre-defined downsampler, these priori hypotheses make the SR methods suffer from a common defect: their models were specialized for a single degradation and lack scalability to handle multiple degradations by a single model
EDSR assumed that the LR image was generated from HR image by bicubic interpolation with no other degradation, we used the EDSR to restore the LR images that were generated from the HR images by bicubic downsampling along with different types of degradations
Summary
As one of the most active research topics, single image superresolution (SISR) has attracted increasingly attention. Examining the existing SR methods that offered optimal working blur kernel, a sharper kernel often leads to the ringing effects and a smoother kernel often leads to blurry output images These phenomena empirically inspires us that the SR model needs robust image prior as an anchor to stabilize the SR results in different Gaussian blurs. In order to go deeper in network structure and achieve better performance, VDSR [23] was proposed using skip connection learning strategy They showed that VDSR could handle multiple SR scales with single CNN model. It achieved good performance, VDSR suffered from high computational budgets because the input images were interpolated to the same spatial size as the output images via bicubic method. SRMD [19] proposed a stretching strategy to integrate the degradation information in the SR network and achieved visually plausible results in reconstructing real LR image
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
