Research On Color Image Denoising Algorithm Based On Improved Generative Adversarial Network In Quaternion Wavelet Domain

As one of the most important media for information transmission,image is widely used in various fields.With the development of information network and intelligent technology,applications such as video monitoring and target tracking have set higher standards for image quality in various scenes of people's lives.However,noise in the image may cause deviation of processing results.Therefore,it is of great research significance to seek a new breakthrough in image denoising technology and improve the quality of image denoising.In recent years,the deep learning based denoising method has achieved good results in gaussian denoising.However,satisfactory results can not be obtained when removing multiple mixed noises or complex blind noises in real scenes.In addition,in the model training,most of the mean square error as a loss function.It can measure the pixel-level approximation between images,but it cannot measure the similarity between image textures.Although the overall visual effect of the denoised image is similar to that of the noiseless image,the denoised image will lose more low-level features and high-frequency information,resulting in smooth texture information.To solve these problems,two different denoising algorithms are proposed in this paper.The main work of this paper is as follows:(1)The loss function of mean square error will result in the loss of high-frequency information in the denoised image and the loss of underlying features.In this paper,an image denoising algorithm(wgan-mv)based on the improved generation adversarial network is proposed.While using the mean square error to measure the overall differences of images,the model adds Perceptual Loss to better measure the Perceptual characteristics between images.Joint discrimination loss(measure local difference)directs the generation network training and learning.Finally,the experimental results show that this method can make the denoised image more similar to the original image in texture details and obtain better visual effects.(2)This paper deals with the removal of complex blind noise in mixed noise or real scene.In this paper,an image denoising algorithm(QG-CNN)based on quaternion wavelet transform and generative adversial network is proposed.Through quaternion wavelet transform,the correlation between the features of each channel of the image is maintained.In the quaternion wavelet domain,the modified generated adversial network(gan-se)is used to implicitly model the features of various mixed noises.While overcoming the difficulty of defining the unknown noise model,the generating network is used to generate noise samples.By expanding and enriching the diversity of sample dataset,the denoising performance and generalization of the learning based model are improved.Finally,experimental results show that the denoising method verifies the accuracy of noise modeling by GAN.A satisfactory visual effect was obtained.In addition,this paper provides a new and effective solution for removing multiple mixed or complex distributed blind noises in images.