Research On Image Superresolution Reconstruction Algorithm Based On Deep Learning

Recently,with the rapid development of information technology and multimedia technology,people have higher demand for the quality of image as one of main media.Currently,the improvement of image spatial resolution is restricted by many factors,such as the density and size of imaging system sensor,network transmission medium,image degradation model itself,etc.In addition,the target motion,light and other interferences in the process of signal acquisition and processing will lead to the decrease of image resolution.Therefore,the super-resolution reconstruction technology based on software and algorithm becomes the main way to solve the above problems under the condition that the hardware of the original system does not change.Super-resolution reconstruction technology aims to convert a given low-resolution image with rough details into a corresponding high-resolution image with better visual quality and high-frequency details,high-resolution image provides important high-frequency detail information for many scene objects,so it has important theoretical significance and application value to study them in depth.With the continuous development of artificial intelligence and computer hardware,Convolutional neural network and generative adversary network have achieved remarkable success in computer vision and image processing.The image super-resolution reconstruction method based on deep learning trains the network model through a large number of sample data sets,learns the complex feature extraction methods,looks for the mapping function relationship between high-resolution and low-resolution images,and finally reconstructs the high-resolution image.This paper mainly studies the optimization and improvement of generative adversarial networks in the field of image super-resolution,and proposes the super-resolution reconstruction method of the dual discriminator generation countermeasure network image and the blind image super-resolution reconstruction method based on CycleGAN.The main research contents and work of this paper are as follows:1.This paper summarizes the research status of deep learning technology and image super-resolution reconstruction technology at home and abroad,and briefly introduces the research background and significance of image super-resolution reconstruction.The image degradation model is briefly introduced,and three types of image super-resolution algorithms based on interpolation,reconstruction and learning are briefly analyzed.2.A super-resolution reconstruction algorithm based on generative adversarial network with dual discriminator is proposed.As a new training method,generative adversarial training is easy to collapse because the generation process is too free and uncontrollable,and the convergence and stability of the network training process are difficult to guarantee,which often causes training process to stop.This paper further studies the generative adversarial network algorithm and improves the network architecture and objective function.In the overall network architecture,another discriminator is added.Kullback Leibler(KL)and reverse KL divergence are combined into a unified objective function to train the two discriminators.By using the complementary statistical characteristics of these two kinds of divergence,the pre estimated density can be decentralized in multi-mode,so as to avoid the model collapse in the reconstruction process,and the stability of model training is improved.3.A blind image super-resolution reconstruction method based on CycleGAN is proposed.Most existing advanced image super-resolution reconstruction methods generally assume that the down-sampled blur kernel is known,such as generating training data and test data by down-sampled high-resolution color images using a fixed bi-cubic blur kernel,but in practical applications,the blur kernel involved in an image is usually complicated,and the low-resolution image generated in this way is not similar to the image captured in reality.In this paper,the CycleGAN model is used for blind image super-resolution tasks.Using the cycle consistency characteristics of CycleGAN,training is performed with unpaired degraded LR and HR images.The experimental results show that the low-resolution image generated by our method is more similar to the real low-resolution image compared with those generated by the bicubic interpolation down sampling method.Compared with many other classical methods,this method can reconstruct the high frequency edge detail texture information more clearly.