摘要
针对目前的稀疏去噪算法分解效率低、去噪效果不理想的问题,提出了一种基于自适应匹配追踪的图像去噪算法。该算法首先通过自适应匹配追踪算法求解稀疏系数,然后利用K奇异值分解算法将字典训练成能够有效反映图像结构特征的自适应字典,最后将稀疏系数与自适应字典相结合来重构图像。在重构过程中,将噪声对应的系数去除,最终达到去噪的效果。算法引入Spike-Slab先验来引导稀疏系数矩阵的稀疏性,并利用两个权重矩阵促使去噪模型更加真实。鉴于字典在稀疏算法中的重要性,将自适应字典与DCT冗余字典、Global字典进行比较。实验结果显示,选择自适应字典的去噪结果比传统字典在峰值信噪比上高出约4.5 dB;与目前6种主流的稀疏去噪方法相比,文中提出的方法在3种评价指标上均有不同程度的提高,其中峰值信噪比平均提高了约0.76~6.24 dB,特征相似度平均提高了约0.012~0.082,结构相似性平均提高了约0.015~0.108。对图像去噪算法进行定性的评价,结果显示所提算法保留了更多的有用信息,视觉效果最佳。实验充分证明了自适应匹配追踪图像去噪算法对图像去噪的有效性和鲁棒性。
Aiming at the problem that the current sparse denoising algorithm has low decomposition efficiency and unsatisfactory denoising effect,an image denoising algorithm based on adaptive matching pursuit was proposed.Firstly,the algorithm uses the adaptive matching pursuit algorithm to solve the sparse coefficients,and then uses the K-means singular value decomposition algorithm to train the dictionary into an adaptive dictionary that can effectively reflect the image structure features.Finally,the ima-ge is reconstructed by combining the sparse coefficient with the adaptive dictionary.During the reconstruction process,the coefficients corresponding to the noise are removed,and finally the denoising effect is achieved.Spike-Slab priori is introduced to guide the sparsity of sparse coefficient matrix,and two weight matrices are used to make the denoising model more realistic.In view of the importance of dictionary in sparse algorithm,this paper compared adaptive dictionary with DCT redundant dictionary and Global dictionary.The experimental results show that the denoising result of adaptive dictionary is about 4.5 dB higher than that of traditional dictionary in terms of peak signal-to-noise ratio(PSNR).The proposed method improves three evaluation indicators in varying degrees compared with the current six main methods of sparse denoising.The PSNR is increased by about 0.76 dB to 6.24 dB,the feature similarity(FSIM)is increased by about 0.012 to 0.082,and the structure similarity(SSIM)is increased by about 0.015 to 0.108 on average.The qualitative evaluation of the image denoising algorithm shows that the proposed algorithm retains more useful information and has the best visual effect.Therefore,the experiment fully proves its effectiveness and robustness.
作者
李桂会
李晋江
范辉
LI Gui-hui;LI Jin-jiang;FAN Hui(School of Computer Science and Technology,Shandong Technology and Business University,Yantai,Shandong 264000,China;Co-innovation Center of Shandong Colleges and Universities:Future Intelligent Computing,Yantai,Shandong 264000,China)
出处
《计算机科学》
CSCD
北大核心
2020年第1期176-185,共10页
Computer Science
基金
国家自然科学基金(61472227,61772319,61602277)~~
