The Research And Implementation On Image Denoising Based On Sparse Representation

The image must be inevitably affected by external environment and becomes noised in the transmission. The results of denoising affect the subsequent operations directly. Recently sparse representation theory based on the Orthogonal Matching Pursuit has been widely used in image denoising and plays a more and more important role. Sparse representation algorithm and the choice of the dictionary are the main research centents. On the basis of the sparse representation and the dictionary learning, the dissertation studied and researched the process of image denoising. The main tasks include the following contents:Firstly, image denoising based on sparse representation usually uses blocks treating thought to preprocess noised image. After preprocessing the noised image, a matrix is produced. But different pixels have different times when they appear in the matrix. The closer to the border, the less times they have while engaging in the denoising process and the least time is only one time. To get the whole image information, all columns in the matrix need to be denoised. This dissertation proposes a new method and restricts a new matrix before using the image blocks treating thought. Then use image blocks treating thought to preprocess image and pick up columns from the produced matrix randomly. Compared to the method used in Journal 27, the outcomes of the experiments demonstrate that the PSNR of this new method decreases about 0.15 dB, but computational time decreases by 12 percentage.Secondly, a trained database is needed while using K-SVD algorithm to train dictionary. The data in database come from the matrix procuded by blocks treating process. The dissertation puts forward an improved method to extract blocks used to train dictionary based on K-SVD algorithm. Classic algorithm extracts the small image blocks from matrix as trained data randomly. To ensure the trained dictionary can describe image better, selected blocks should contain more original information and the number of the selected blocks shoule be bigger. There is more information overlapping among those blocks. The dissertation puts forward a new method to extract trained data. Extract blocks whose column number j satisfy the condition j =8k +1 and k is an integer not less than zero as trained data. Compared to the classic K-SVD algorithm, the results of the experiments show that the PSNR of the method which improves trained database decreases only 0.1dB, but the working time of the program has good performance, decreases by almost 13 percentage.Thirdly, in K-SVD algorithm, the absolute value of cosine similarity is regarded as the basis for judging the similarity of two atoms. While the absolute value of similarity degree is higher than 0.99, the dictionary is thought to be redundant and use one image block to replace one atom. This dissertation regard two-norm square of the difference between two atoms as the judgement. When the vectors meet the criteria or the opposite vectors of,m nd d belong to dictionary and the first elements of,m nd d have consistent symbols), we regard the dictionary as redundant and use the image blocks to replace the redundant atoms. Simulation results show that this method almost has the same PSNR as the method which uses absolute value of consine similarity and shorter time.