基于清晰图像先验知识的盲复原算法

图像采集系统在获取图像时,受到各种干扰,使图像变模糊并含有噪声。为此,提出了一种基于清晰图像先验知识的图像盲复原算法。首先通过统计分析大量的清晰图像梯度域的概率分布特性,将图像划分为具有较多复杂纹理的区域和平坦区域,拟合概率分布函数,得到清晰图像的先验知识,并加入图像局部约束条件,即复原图像平坦区域与退化图像的平坦区域的概率分布变化较小,避免产生振铃现象。然后,建立图像噪声模型,将图像噪声划分为高斯噪声和均匀分布噪声,防止退化图像中的灰度过饱和像素点在复原过程中产生异常值。最后,利用最大后验概率模型,构造代价函数,将图像盲复原问题转变为求代价函数最小化问题,并用最大期望算法和快速迭代收缩算法求解代价函数,恢复出清晰图像。实验结果表明,该算法能有效恢复图像细节,锐化图像边缘,抑制噪声,避免产生振铃效应。

Image blind restoration using priors of sharp images

The imaging systems are affected by various interferences, causing image blur and noise. In order to solve the problem, a new blind restoration method was proposed to recover the images using knowledge of sharp images. Firstly, the statistics of gradient magnitude values of sharp images were analyzed, and the images were divided into two parts, the highly textured region and the flat region. Fitting the probability distribution functions, the priors of sharp images were obtained, and the constraint was injected which the probability distribution of the flat region was little affected by blur kernel, aiming to avoid the image ringing. Then, the image noise model was established and divided into two parts, the Gauss noise and random noise, with propose of keeping off outliers caused by the over-saturated pixels. Finally, the maximum posterior probability was utilized to construct cost function, and emploied the expectation-maximization algorithm and iterative shrinkage algorithm were emploied to solve the cost function to recover image. The experiments show that it archive in recovering much image details, sharpening edges and avoid artifacts.