Research On Hazy Image Restoration Algorithm Based On Generative Adversarial Network

As the basis of perceiving and recognizing the world,image is an important means for humans to obtain information,express information and transmit information.The image-based information interaction usually involves the process of image acquisition,compression,storage and transmission.In these processes,some external factors inevitably have effect on the image quality,which in turn affects the sufficiency and accuracy of image information expression.With the rapid development of big data and artificial intelligence technology,the demand for images is becoming more prominent.For this reason,digital image processing technology plays a very important role in people's life.As an important branch of digital image processing,image restoration aims to simulate the image degradation process by constructing a mathematical model,and restore the real image from the original degraded image by inverting the mathematical model.At present,image restoration technology is widely used in many low-level visual tasks such as image denoising,image dehazing,image deraining,and image super-resolution.However,there are many reasons for image degradation and their properties are different.Therefore,there is still no unified and efficient image restoration algorithm.Recently,the rapid development of deep learning technology has injected new vitality into traditional image restoration algorithms,and it may be possible to find an efficient and wide-ranging image restoration algorithm.Here,we take image dehazing as an example,and apply the generative adversarial network in deep learning to the hazy image restoration algorithm.The main contents are as follows:(1)The problem of inaccurate transmission map estimation has always existed in traditional image dehazing algorithms.In order to solve this problem,in this chapter,a transmission map prediction network is proposed.The network is based on the CGAN(Conditional Generative Adversarial Networks)framework,and its generator is a multi-scale fully convolutional DenseNet(FC-DenseNet)transmission map prediction network.The network mainly includes three parts: multi-scale information extraction,transmission map feature extraction and fusion,and transmission map prediction module.In order to ensure the accuracy of the transmission image estimation,the network uses the mean square error function as its loss function,supplemented by WGAN(Wasserstein Generative Adversarial Networks)to optimize its generated results.The experimental results show that the method is more accurate in predicting the transmission image,and has more advantages in dehazing effect than other comparison algorithms.(2)The traditional image dehazing algorithm has inaccurate intermediate value estimation and formula calculation.In this chapter,an end-to-end image dehazing method based on WGAN is proposed to solve this problem.First,the above-mentioned multi-scale FC-DenseNet transmission map prediction network is used to improve the accuracy of transmission map estimation;second,a shallow residual dehazing network is introduced,which can effectively avoid subsequent calculation errors caused by inaccurate transmission image prediction by sharing image convolution features with the transmission image prediction network;then,the two networks are jointly optimized by constructing a multi-task loss function;finally,in order to generate a realistic image,WGANs are used to fine-tune the generated transmission maps and dehazing images.A large number of experiments show that the algorithm is better in detail processing and image texture restoration,and it is superior to other comparison algorithms in subjective and objective evaluation.(3)The existing image dehazing algorithms rely heavily on the accurate estimation of intermediate variables.Due to the complexity of the real scene,it is more difficult to produce the corresponding transmission image label.To this end,an end-to-end image dehazing method based on WGAN is proposed.First,to simplify the network model,FC-VoVNet embedded with the multi-scale pooling module is used to fully learn the features of hazy in the image;second,the residual learning idea is adopted to directly learn the features of the clear image from the degraded image to achieve end-to-end dehazing;finally,the mean square error and perceptual structural error function are used as the loss function of the model to ensure the image structure and content information,and WGAN is used to finely optimize the generated results to produce clear and realistic clear images.Experimental results show that the algorithm is superior in dehazing effect.And the method can be applied to other image restoration fields,generalization performance is better.