Research Of BP Neural Network For Image Restoration Based On Differential Evolution Optimization

Blurred images are inevitably caused by noise,distortion and other factors in the process of image acquisition and transmission.But high-quality images are eagerly demanded in practical application,and image restoration technique came into being.According to blurring mechanisms,blurred images are generally divided into three types including noise-blurred images,motion-blurred images and defocus-blurred images.And noise-blurred images includes Gaussian blurred images,Uniform blurred images,Salt&pepper blurred images,etc.Different types of blurred images have different mechanisms of degradation.In other words,Point Spread Functions(PSF)are different.In addition,insufficient prior information for blurred images and unknown PSF are the challenges that image restoration have to face.According to knowing the PSF or not,the image restoration algorithm can be divided into non-blind restoration algorithm and blind restoration algorithm.As blind restoration algorithm,BP neural network has received extensive attention;however it is sensitive to initial value,convergences slowly and is easy to fall into local optimal solution.In this paper,the improved method is proposed to optimize the BP neural network and used it for images restoration.The contributions of this paper are mainly as follows:(1)In the sight of the research on the model of image degradation,this paper describes the basic theory of image restoration and analyzes the imaging mechanism and PSF of common blurred images.At the same time,some classic algorithms are studied,and the evaluation standard of image quality is introduced.(2)To improve the deficiencies of BP neural network for image restoration,this paper proposes random scaling-differential evolution(RSDE)in the frame of differential evolution(DE)and combines RSDE algorithm and BP neural network to restore images.The RSDE algorithm optimizes the fitness function of the RSDE-BP algorithm by adjusting the parameters of the BP neural network so as to catch a better performance.The experimental results show that the proposed algorithm has a good restoration performance in the restoration of Gaussian blurred images.(3)To expand the application of BP neural network for image restoration,an improved sliding window is used to sample the image,and the Levenberg-Marquardt algorithm(LM algorithm)is used to combine with RSDE algorithm.Besides,a new method is proposed for image restoration of BP neural network based on RSDE-LM optimization in this paper.A mapping model between the blurred image and the clear one is established by training the BP network.Then it is utilized to restore the blurred image.The experimental results show that the proposed algorithm has a good restoration performance in the restoration of noise-blurred images,motion-blurred images and defocus-blurred images.What's more,the restoration of real-world images can also meet the restoration requirements.On the basis of BP neural network,two kinds of algorithm are proposed for image restoration of BP neural network based on improved differential evolution.The proposed algorithm not only provides a solution for image restoration of different types of blurred kernels,but also solves the problem of real-world images restoration.