摘要： 稀疏表示因其具有稀疏性、特征保持性等一些特点而被广泛应用于图像处理等领域，为解决图像处理中的去噪问题，提出一种基于图像特征稀疏表示的贝叶斯去噪模型。利用K-means和主成分分析方法计算已分割图像块对应字典的矩阵系数，采用正则化约束条件，迭代计算获取的图像字典与原始图像字典之间的差距，优化噪声图片稀疏特征表示的字典，直到达到优化条件。实验结果表明，与传统的离散余弦变换去噪模型相比，该模型的峰值信噪比较高，随着噪声的不断提高，与噪声图像峰值信噪比的差距也越来越大，且图像失真较少。

关键词: 图像去噪, 字典学习, 贝叶斯模型, 稀疏表示, 正则化, 高斯噪声

Abstract: For the sparse characteristic and maintaining features characteristic, the sparse representation is widely used in image processing. To solve the problem of image denoising in the area of image processing, this paper proposes a new Bayesian denoising model based on image feature sparse representation. The model uses the K-means and Principal Component Analysis(PCA) method to obtain the coefficients of dictionary for sparse representation solutions of image patches. The coefficients solutions are used to train the dictionary with regularized optimization. The alternating minimizations are kept between above two steps until the difference between the image dictionary and the source image dictionary satisfied a convergence criterion. It restores the denoising image under the MAP model with that dictionary. Experimental results show that the higher Peak Signal to Noise Ratio(PSNR) value than the source noised images with the increase of imposed noise into clean images, comparing to the initialization with Discrete Cosine Transform(DCT).

Key words: image denoising, dictionary learning, Bayesian model, sparse representation, regularization, Gauss noise

中图分类号: 

• TP391