Research On Single Image Super-resolution In Complex Situations

Single image super-resolution(SISR)aims to recover a high-resolution(HR)image from a single low-resolution(LR)image.Nowadays,this technology has been applied in many key areas,such as aerospace,geographic information,biomedicine and security monitoring.This paper researches SISR in complex situations.Complex situations mean that all factors possibly causing image quality degradation in the real world,such as noise,blurring,and pixel loss,need to be considered comprehensively.This research is much beneficial to actual demands of production and living,and practical in terms of the research significance and value.Firstly,this paper introduces the research background and significance,the development process and the status at home and abroad of SISR.Afterward,the causes of image degradation,probability density functions of image noises,image pyramids,Gaussian blur and down-sampling of images are analyzed,which can be further summarized into an image noise model and an image degradation model.Finally,the strategy of removing salt and pepper noise(SPN)before image reconstruction is proposed.For the removal of SPN in images,this paper proposes an adaptive weighted kriging interpolation filter(AWKIF),which employs inverse filtering-radiuses to estimate the calculation weight of processing windows.This method not only suppress the effect of pixel aggregation,but also enables pixels to be calculated according to the principle of proximity.The effectiveness of the proposed algorithm is demonstrated by the experiment on the databases of Kodak and Caltech101.In order to achieve super-resolution of degraded low-resolution images,this paper proposes a method based on deep dynamic convolutional neural networks(DDCNNs),which firstly establishes the mapping relation of image degradation via an extraction strategy of image details,then learns this mapping relation and reconstructs the details of high-resolution images using dynamic convolutional neural networks.In the experiment,the performance of the proposed method is tested on the standard databases(Set5,Set14 and BSD100)and some randomly screened scene images from the real world.The experimental results show that the proposed method features fast learning and high accuracy.In conclusion,this paper studies SISR in complex situations.Based on the characters of image quality degradation in complex situations,this paper proposes solutions from two perspectives: the removal of SPN and image reconstruction via deep dynamic convolutional neural networks.The proposed AWKIF and DDCNNs are able to suppress fixed value and random noise,blurring,pixel loss in images,so they are effective to solve the problem of SISR in complex conditions.