光学 精密工程, 2013, 21 (7): 1906, 网络出版: 2013-08-05
保留结构特征的稀疏性正则化图像修复
Feature retained image inpainting based on sparsity regularization
图像修复 剪切波变换 稀疏性正则化 增广Lagrange函数 image inpainting shearlet transform sparse regularization augmented Lagrange function
摘要
以压缩传感和稀疏表示为理论依据, 提出了一种基于剪切波变换的稀疏性正则化的图像修复模型, 以便更好地保留图像的结构特征。该模型用剪切波作为图像的稀疏表示, 以稀疏性作为正则化项; 同时基于变量分裂法, 采用增广Lagrange优化方法求解最优化问题。另外, 通过交替最小化方式来降低计算复杂性。从峰值信噪比(PSNR)、结构相似度(SSIM)、收敛速度和视觉效果等4个方面验证了算法的有效性。结果显示:利用本文算法修复图像的质量明显优于其他算法, 获得了更优的PSNR和SSIM值。新的模型无论是在客观还是视觉主观方面都具有更好的性能, 同时算法具有更快的收敛速度。得到的结果表明本文算法能够更好地修复图像, 获得较好的视觉效果。
Abstract
By taking compressed sensing and sparse representation as theoretical bases, a sparse regularization image inpainting model based on shear wave transform is proposed to reserve the structure characteristics of an image. The model uses shear wave as sparse representation and sparse as a regularization item.Meanwhile, based on variable splitting method, it uses augmented Lagrange method to solve the optimization model. Furthermore, it reduces the complexity of the calculation through alternating direction method of multipliers. The availability of the algorithm is verified by Peak Signal to Noise Radio(PSNR), Structural Similarity Index (SSIM), convergence speed and visual effect. The results indicate that the image inpainting quality by proposed algorithm is better than that by other algorithms, and more optimal PSNR and SSIM can be obtained. The new model has more better performance whether for objective or for visual passitive, moreover, it shows a far quicker convergence rate. It concludes that the algorithm can inpaint images effectively and obtain a better visual effect.
邓承志, 刘娟娟, 汪胜前, 朱华生. 保留结构特征的稀疏性正则化图像修复[J]. 光学 精密工程, 2013, 21(7): 1906. DENG Cheng-zhi, LIU Juan-juan, WANG Sheng-qian, ZHU Hua-sheng. Feature retained image inpainting based on sparsity regularization[J]. Optics and Precision Engineering, 2013, 21(7): 1906.