Color Image Denoise Method With Generative Adversarial Nets

In the process of image acquisition and transmission,lots of image noise will be generated.When the image is acquired,the photoelectric conversion of the device will introduce noise.And when the image is transmitted,the channel noise will also interfere with the image,so the image quality will tend to decrease.The image noise will make the information contained in the image indeterminate,so that people can not recognize and understand the images well.In the field of computer vision,when image processing operations such as recognition and segmentation are performed,image noise causes serious deviations in processing results.In the military and medical fields,the errors caused by such deviations will bring huge losses.Therefore,image denoising has important research significance and it is a hot topic in computer vision.In this paper,we use the deep learning method for image denoising.In view of the generative adversarial networks have achieved excellent performance in image super-resolution,image deblurring and image scene conversion tasks.It is fully proved that using generative adversarial networks to do pixel-level operation will better recovery the texture details of images.Therefore,this paper adopts generative adversarial networks to denoise images.A novel generator is built as a denoising network.In order to make denoised images preserve more texture details of the original images,noisy images are transformed into the feature domain through a neural network when denoising.In the network,multi-scale features are extracted using different-sized convolutional kernels,and the extracted noisy features are denoised and filtered,and then merged to denoised images.By learning a large number of samples in image database,the network can extract the texture features of the image details,and use the rich information stored in multi-scale features to preserve texture details in the images.In the network,multiple skip-connection structures are used to accelerate the convergence of the network while retaining detailed information.We use discriminator network and generator network for adversarial training.Training discriminator enables it to discriminate the difference between clean images and generated denoised images,and we feed the difference back to generator network,so that make the network generate denoising results which are more similar to the clean images and preserve more image textures.In adversarial training,a novel loss function is proposed in this paper to guide the training of the generator,by adding a new loss term which is used to represent the distance between the data distribution of the clean image and the denoised image,so that the network optimization results are closer to the real images.In order to ensure the correctness of discriminator network,we need to train the generative adversarial networks by using noisy images with fixed noise level.For images with different level noises,different network parameters should be used to denoise them.In order to adaptively select the network parameters,this paper builds a convolutional neural network to estimate the noise level,and loads corresponding network parameters according to estimated results,so as to complete image denoising process.The proposed method is compared with other image denoising methods.The superiority of this method is demonstrated by quantitative objective evaluation indicators and qualitative subjective visual evaluation.