多正则化混合约束的模糊图像盲复原方法
doi: 10.11999/JEIT140949
Multi-regularization Hybrid Constraints Method for Blind Image Restoration
-
摘要: 图像复原是一个长期的且极具挑战性的逆问题。为了实现模糊图像的盲复原,该文提出一种多正则化混合约束的模糊图像盲复原方法。首先,运用一种图像的局部结构提取策略(Local Structure Extraction Scheme, LSES)将图像中的大尺度图像边缘准确地提取出来。然后,在模糊核(Blur Kernel, BK)的估计阶段,将提取的大尺度图像边缘与前期研究中所提出的一种结合稀疏性和平滑特性的双重正则化约束模型相结合,实现模糊核更加准确的估计。在图像的复原阶段,为了得到高质量的复原图像,提出一种结合全变差(Total Variation, TV)模型和Shock滤波器不变特性的多正则化约束模型,从而实现模糊图像的清晰化复原。最后,通过半二次性的变量分裂策略对提出的模型进行最优化求解,能够在准确地估计出BK的同时得到高质量的复原图像。在人造的模糊图像和真实的模糊图像中进行了大量的实验,证明了所提方法的有效性,且与近几年的一些极具代表性的模糊图像盲复原方法相比,不仅主观视觉效果得到了显著的增强,而且客观评价指标也得到了明显的改进。
-
关键词:
- 图像盲复原 /
- 多正则化混合约束 /
- 局部结构提取策略 /
- 全变差(TV)模型 /
- Shock滤波器不变特性
Abstract: Image restoration is a long-standing and challenging inverse issue. In order to recover an image from its blurry version blindly, a multi-regularization hybrid constraints method is proposed. First, the large scale edges are extracted from the image with a Local Structure Extraction Scheme (LSES). Then, in the Blur Kernel (BK) estimation step, the extracted large scale edges are used for BK estimation, and a sparsity and smoothness dual-regularization constraints model proposed in the previous study, is also employed for estimating BK more accurately. In the image restoration step, a multi-regularization constraints model, which combines the Total Variation (TV) model and Shock filtering invariance, is proposed for obtaining high-quality restoration image. Finally, in order to exactly estimate the BK and simultaneously obtain high-quality restoration image, the proposed models are addressed with a half-quadratic variables splitting scheme. A large number of experiments are performed on both synthetic blurred images and real-life blurred images. The experimental results demonstrate the effectiveness of the proposed method, while in comparison with several recent representative image blind restoration methods, not only the subjective vision, but also the objective numerical measurement has obvious improvement.
计量
- 文章访问数: 1763
- HTML全文浏览量: 120
- PDF下载量: 857
- 被引次数: 0