Research Of The Image Restoration Algorithm Based On Generative Adversarial Networks

In recent years,with the development of computer vision and artificial intelligence,image restoration technology based on deep learning has been widely used in medicine,military,monitoring system,photography and film industry.Image restoration refers to complementing the missing area and removing image noise using neighborhood image information and image structure information.So far,many image restoration algorithms have been proposed.Although these algorithms can repair the damaged image to some extent,there are still some problems such as the inability to retain the texture information of the original image.In order to solve these problems,we make a study on the image inpainting algorithm and the image denosing algorithm based on Generative Adversarial Networks.The innovative achievements are summarized as follows:(1)A stepwise image inpainting algorithm based on Generative Adversarial Networks is proposed.In the algorithm,two denoising models are used to repair the image.The pre-inpainting model is used to preliminarily inpainting the image and restore the low-dimensional structure information of the image.And the enhanced-inpainting model restores high-dimensional texture information of image on the basis of the pre-inpainting model.The experimental results fully demonstrate the effectiveness and feasibility of the proposed algorithm in image inpainting.(2)In order to solve the problem that existing image denoising algorithms lose a lot of image texture information while denoising,an image denoising algorithm based on asymmetric Generative Adversarial Network is proposed.In order to balance the performance of image texture information preserve and denoising,a down-sampling layer is added between the denoising model and the discriminanting model.On this basis,the input image size of the discriminanting model is reduced and the training iteration of the discriminanting model are increased.It creates an asymmetry between the denoising model and the discriminanting model in image size and training iterations.Experiments show that the proposed image denoising algorithm has good performance in evaluation indicator and visual perception.(3)By simplifying and improving the proposed asymmetric Generative Adversarial Network model,a modified image blind denoising algorithm based on asymmetric Generative Adversarial Network is proposed.The denoising model is simplified compared to the previous model,and a multi-scale down-sampling layer is added to extract more dimensional image information for denoising.In addition to image blind denoising,the model is also applied to image deraining.The experimental results fully demonstrate that the proposed algorithm has improved both in visual effect and evaluation indicator compared with the current image denoising algorithm.