Research On Image Denoising Based On Generative Adversarial Networks

With the rapid development of modern computer technology,more and more images from natural scenes have been collected and further analyzed in military fields,scientific research as well as computer vision fields.However,the definition of those collected images may be influenced by rainy and foggy weather,thus would reduce the visibility of images,which would not only affect the visual quality,but also affect the relating image processing research,especially in some scientific fields such as automatic driving,video surveillance,etc.Therefore,some practical methods are urged to be found to realize image denoising.An improved image rain and fog removal algorithm is proposed based on the existing generative adversarial networks model in this paper.By combining the idea of image style conversion with image denoising process,the loss function of existing generation model is improved,and a more universal image denoising algorithm is proposed.On the problem of image rain removal,a new image rain removal algorithm was proposed by adding vgg-16 module to the original network structure to extract the perceptual loss and improve the loss function.In the problem of image defogging,we improved the Pix2 pix HD model.By using the multi-scale generation and discrimination network to train the data set,super resolution processing has been realized,which ensuring the definition of generated images.Meanwhile,a tone constraint item is added into the network to restrict the training process.By improving the contrast of the generated defogging image,no obvious color difference may appear between it and the target image,so as to improve the quality of the generated image.Compared with some existing ideal image denoising methods,the main contributions of this paper are as follows:(1)The image denoising model improved in this paper regards the image denoising process as the style conversion problem of images in the same scene.Training the network with paired pictures does not require a priori model setting,furthermore,the training speed of the network is faster and the model may converge more quickly.(2)The focus of raindrop removal problem is to ensure the definition of generated images and the restoration of the rain occlusion content.Based on the existing model,we add the VGG-16 module to the original generative network to extract perceptual loss as a new component of the loss function.The new network model with added perceptual loss can better restore the detailed information at feature level,which further ensuring the consistency of the picture content before and after the process of rain removal.(3)In the experiment of image defogging,the key is to remove the fog through the defogging operation.While ensuring the definition of the generated images,it is also necessary to ensure that the contrast of the picture content is consistent with the target pictures,so that the picture will not have a large color shift.This step is mainly achieved by improving the picture contrast.The image defogging model proposed in this dissertation is improved based on the Pix2 pix HD network model.The multi-scale generator ensures that more local details can be learned during the training process.Combined with multi-scale discriminator,the entire model generates images with higher resolution.While realizing the high definition processing of the generated images,the generated picture will not have color shift by adding the hue constraint during the training process,so that the experimental results have a higher fidelity.