Research Of Single Hazed Image Denoising Based On Variational Method

With the development of technology, various kinds of electronic and digital products are increasingly common in daily life. Images have already become the critical way to obtain the information from outside. However, noise would be introduced unavoidably in the procedure of obtaining, transferring and restoring an image, thus greatly decreases its quality. On the other hand, noise will make the subsequent processing, for instance, image segmentation, edge extraction and object recognition, more difficult. Therefore, image denoising becomes an indispensable part of digital image processing. The goal is to not only maintain the completeness of original information, but also remove the useless signal thus improve the image quality as much as possible. This is an important pre-step to analyze and understand images.People have been studied and explored a lot in this area for a long time. Image denoising methods based on partial differential equation become the hot point gradually. Total variation is also used commonly in the practical problem. This paper will discuss single haze image denoising in detail.The classical image denoising model is the Rudin-Osher-Fatemi model. There are also similar ones, such as H1 regularization, primal-dual. These models take advantage of the gradient information to remove noise from an image. However, outdoor images are different because of their own unique characters. The noise in this kind of images is introduced by two ways:one is because of the outside environment, for example, atmosphere, thick smoke and haze; the other one is related with the brightness of light. When the light environment is too dark, the ISO of the camera will be very high thus produce the different kind of noise in the image. To solve this problem, this paper combines dark channel prior theory and soft matting refinement methods to come up with a novel variation model which creates a new energy functional. This method can remove both of the two different types of noise without any other additional equipment and information.In addition, in order to accelerate the convergence, the Split Bregman iteration method will substitute the traditional gradient descent to solve the problem. This greatly decreases the time of computation and complexity of our algorithm as well. PSNR and MSE methods are used to measure the results at the end of this paper. Related analysis and discussion will also be presented in detail.