Research On Blind Super-Resolution Algorithms Based On Deep Learning

Most super resolution(SR)methods assume that the low resolution(LR)image is obtained from the high resolution(HR)image by a fixed degradation method(such as bicubic interpolation downsampling).However,the degradation process of LR images is various in the real world.When the degradation process of the LR images differs greatly from that assumed by the SR model,SR performance significantly deteriorates.Therefore,blind SR method is a research trend.While SR models need an accurate kernel estimation method to achieve better performance of blind SR.Firstly,this thesis proposes a kernel estimation method based on the Fourier transform and convolutional neural network(FTKCN).The Fourier transform is used to generate the frequency spectrum from the LR image.The convolutional neural network is used to extract the blur kernel information from the frequency spectrum,so as to get the blur kernel.FTKCN converges faster in training,and has higher accuracy and better stability.Then,the proposed kernel estimation method is combined with the three SR methods of ZSSR,MZSR,and SRMD.The kernel generated by FTKCN is used to generate training data or input to the network for modeling to help SR tasks.Experiments show that FTKCN helps to improve the blind SR performance of these three SR methods.Finally,this thesis introduces the structure of dense connections on the basis of SRMD,and proposes a super resolution network with dense blocks and multidegradations(DenseMD).DenseMD makes full use of the degradation information,and only requires a small amount of parameters to obtain better performance in the task of blind SR.In addition,combining DenseMD with FTKCN can further improve the performance of blind SR.