Image Super-Resolution Reconstruction Based On Generative Adversarial Networks

Image plays an important role in human social development,information storage and cognitive learning.However,due to the limitations of acquisition hardware equipment and network transmission,image resolution is not high and the use of general effect.The more detailed information a high-resolution image has,the more helpful it can be.Therefore,the image super-resolution reconstruction technology arises at the historic moment.Image super resolution reconstruction refers to the use of software to improve image resolution without upgrading the hardware,starting from the computer digital technology.At present,the technology has made extensive development in public security,medicine,aviation,art and other aspects.Therefore,on the basis of existing research,this article obtains from the method based on learning,primarily for emergent against network in the process of super resolution image reconstruction,with the deepening of the network structure,network appear gradient disappear or explosion,model degradation problems,and reconstruction of image detail resume the phenomenon such as imbalance result is bad,the color deviation,relevant the oretical research and model improvement are carried out,and a new model structure is constructed.The main research work of this thesis can be summarized into the following two aspects:On the basis of SRGAN network,use a Dense Net block with dense connection mode instead of the original Res Net block,and a Dense-SRGAN network based on SRGAN was proposed.The significance of batch standardization layer in restoring superresolution images was discussed,and then the optimal design of SRGAN network was carried out.This thesis puts forward a kind of based on low-level features generated type against network LF-SRGAN super resolution reconstruction technology,the network is integrated with the feature pyramid structure on the basis of dense connection modules,and in view of the reconstruction image characteristics,adjust the relevant network parameters,constitute the main for reconstruction of lower resolution image low-level information generated against network,the low level features of the image can be further reused.To sum up,the proposed low-level feature Super-resolution generative countermeasures network LF-SRGAN based on SRGAN has higher objective evaluation indexes and better subjective visual effects,which can improve the image quality.Comepared with the traditional Super-resolution model,the algorithm proposed in this thesis has achieved some improvement.