Research On Image Restoration Method Based On Convolutional Neural Network

With the continuous development of virtual reality technology and mobile Internet,image plays an important role in the process of gaining and passing information for people.In everyday life,our vision is flooded by all kinds of images and video.Since the human eye is sensitive to image definition,low resolution images could bring uncomfortable feelings to the viewers.In order to meet the demands of the people,the display technology is developed rapidly.However,there are too many unforeseeable elements in real life which could bring about image distortion,such as being out of focus and camera shake.So it's very important to restore the original image using distorted image.This type of technology is called image restoration,which is relatively popular in the field of research at present.In this paper,we take advantage of convolutional neural network(CNN)in image restoration,and investigate two hot issues of image restoration: image deblurring and super-resolution reconstruction.This thesis contains following contents:(1)An image deblurring model based on CNN is presented.In comparison with the traditional image deblurring algorithm,this model can avoid the dependence on apriori knowledge of blurred image.On the condition that Point Spread Function(PSF)is unknown,the network gets the inherent relation between input image blurred and output clear image,and the image deblurring is implemented.When selecting the network parameters,a trade-off between time and performance is made by experiment.Then the best parameters are applied to the model.Experiment result shows that the model based on CNN has higher precision and better performance than traditional image deblurring algorithm.(2)An image deblurring method based on hybrid neural network is presented.The hybrid neural network is composed of CNN and BPNN,which realizing image restoration step by step.Firstly,though training CNN,the effective perception features from blurred image is extracted.Secondly,the features are using as the inputs to train the BPNN,in order to realize image restoration.Experiment result shows that the restoration effect is better than existing methods on small fuzzy kernel scale.However,when the fuzzy kernel scale exceeds 23x23,the restoration effect is very poor.(3)An improved image super-resolution reconstruction model based on CNN is presented.The image super-resolution reconstruction model based on CNN contains three convolutional layers,which are used as image feature extracting,nonlinear mapping and reconstruction.In this paper,we present an improved image super-resolution reconstruction technology by adding network layers,changing the number of filters in convolutional layers and changing the scale of filters in convolutional layers.The improved CNN contains four convolutional layers and one pooling layer.The max,medium and min pool structure is using by the poolinglayer,which is help to effectively extract the perception features in image and improve training efficiency.Experiment result shows that the model can exactly approximate real higher resolution images.