Studies On Image Patch Based Denoising Algorithm

The most direct way for human to perceive and know the outside world is to see from the images, which convey important information. A clear image without pollution can ensure the authenticity of the information. However, in the process of image imaging or transferring, noise can be inevitably introduced, which leading to the information distortion. So, we need to restore the clear image from the noised image. This operation is also called image denoising. Image denoising is widely used in many fields, such as medical image processing, meteorological remote sensing, biometrics, security monitoring, target tracking, etc. Therefore, the study of image denoising is of great significance and broad application prospects.In recent years, with the advantages of low computational complexity and better denoising performance, many researchers devote themselves to patch-based denoising methods. This thesis makes attempts to improve some excellent patch-based image denoising algorithms. The main jobs of this thesis are described as follows.(1) The improvement of Expected Patch Log Likelihood (EPLL) algorithm. The EPLL algorithm does not incorporate the internal structures of the local image, which uses the fixed parameters of Gaussian mixture models. To solve the problem, this thesis tries to adaptively update the parameters of Gaussian mixture models. We learn the parameters of the patches of the denoised image at each iteration. Then, we can use the new Gaussian mixture models to denoise the image. Repeat the operation until a satisfactory result is got.The experimental results show that the improved algorithm is able to retain the image structure information and can improve the denoising performance compared to the state-of-art denosing algorithms.(2) The improvement of nonlocally centralized sparse representation (NCSR) algorithm. The NCSR algorithm solves the problem of image restoration based on sparse representation. It restores the image by reducing the sparse coding noises. However, as its sparse coefficients of the original image are estimated by calculating a certain number of weighted average of similar image patches, the accuracy is not enough. Moreover, an estimation bias inevitably happens because the noises are correlated in the intermediate process. This thesis uses the generalized non-local means to achieve a more accurate estimation. The experimental results show that the improved method can achieve an excellent restoration performance and outperform other algorithms.