Image Inpainting Based On Dynamic Parameter Selection Multi-exposure Fusion And Generation Adversarial Network

Image is an important means of carrying information,which is indispensable in human daily life.However,due to illumination,color aberration,environmental noise interference and other reasons in the process of image acquisition,the finally collected image cannot be guaranteed to be clear and intuitive every time.In addition,in the process of image transfer,due to compression technology and format reasons,the quality of the obtained image may be damaged to a certain extent.Image restoration technology based on neural networks has developed rapidly before,and the emergence of Generative adversarial network(GAN)provides a new idea for image restoration.The image of data set is often attacked by noise in the process of acquiring,transmitting and processing.The missing area of the restored image is discontinuous with the non-missing area,and there is some trace of repair.The complex network structure increases the time complexity and space complexity of the model.In this study,a multi-exposure fusion method based on dynamic parameter selection is proposed to preprocess the data set,and a Generative adversarial network is constructed to repair the missing image.There are three specific points:1.Digital image is often attacked by noise in the process of acquisition,transmission and processing,which will seriously affect a series of subsequent operations such as feature extraction.A multi-exposure fusion method based on dynamic parameter selection is proposed to preprocess the acquired images.2.The missing area of the restored image is discontinuous with the non-missing area,and there is a trace of repair.A new generative adversarial network is used to repair the image.Generator part convolution layer adopts extended convolution to better extract the feature information of the original image.The discriminant network adopts global and local discrimination,and three loss functions are designed to optimize the generated adversarial network,so that the restored image is more real and natural.3.As complex network structures increase the time and space complexity of the model,Region Normalization(RN)is used to overcome the problem of poor network training in order to help network training better.RN divides spatial pixels into different regions according to the input mask,and calculates the mean value and variance of each region for normalization.Experimental results show that our method has good performance in image exposure,color saturation,clarity and local details in data set preprocessing,and the objective indexes of data set damage repair are better than other methods in comparative experiments.