Research Of K-SVD Algorithm In Image Denoising

In reality, images tend to be influenced by noise from equipments or the external environment, resulting in the severe decrease of image quality. And such noise is problematic to the image interpretation, making image denoising a critical step in the image preprocessing phase, such as image feature extraction, segmentation, recognition and so on.The main work of this paper is on the K-SVD algorithm which is applied into image denoising. According to the drawbacks of K-SVD, we carry out the research, and propose three new denoising algorithms, and demonstrate their effectiveness from theoretical and experimental aspects. The main contributions of this paper include the following three aspects:(1) The Orthogonal Matching Pursuit (OMP) algorithm and Singular Value Decomposition (SVD) algorithm are used in the K-SVD algorithm. When the image size is too large, the effectiveness of OMP algorithm becomes limited, besides, the SVD algorithm is not only time consuming, but also occupys much space, which usually cause "out of memory" phenomenon. To solve these problems, this paper proposes a sparse K-SVD algorithm which is based on the chelesky decomposing and approximated SVD for natural image denoising. Finally the effectiveness of K-SVD is improved largely and the "out of memory" phenomenon is solved too.(2) The original aim of K-SVD algorithm is to process natural images whose noise is additive, while the SAR image speckles belong to multiply noise, the oversmooth phenomenon will be caused by applying the K-SVD algorithm to SAR images directly. To overcome this shortage, this paper studies the statistic properties of SAR images, and then uses maximum likelihood estimation, finally deserves a new K-SVD object which is suited for SAR images despeckling. Its rationality is proved theoretically, we call it SARK-SVD algorithm in this paper. Not using the log transformation, the SARK-SVD algorithm not only effectively keeps the original radiation characteristic, but also eliminates the oversmooth phenomenon, such as the blur of edges, objects and texture features etc, which is caused by directly applying the traditional K-SVD to SAR images despeckling.(3) Different surface features owning different backscattering coefficients in the SAR image, so in order to further improve the SAR_K-SVD despeckling result, this paper classifies the SAR image pixels, and then introduces multilayer category in which the sparse representations are got by using different dictionaries whose sizes are different to suppress the speckles, we call it the multilayer SARK-SVD algorithm in this paper. Compared with the SARK-SVD algorithm, the despeckling results which are got by multilayer SAR_K-SVD algorithm are improved obviously.