基于变分狄利克雷模糊核估计的行车记录盲图像复原

Vehicular Image Restoration Based on Variational Dirichlet Blur Kernel Estimation

针对行车记录模糊图像的模糊核不能准确估计的问题,提出一种变分狄利克雷分布的模糊核估计方法,并利用改进的增广拉格朗日算法实现有效的图像复原。采用梯度投影法求解优化问题,提取图像边缘的精确方向,用狄利克雷分布替代模糊后验估计消除图像噪声,减少额外约束;以超拉普拉斯先验分布建模,结合ALM实现行车记录盲图像复原。实验结果表明,通过多尺度的模糊核估计,能有效估计模糊核并消除行车记录图像的噪声,恢复行车记录图像的纹理细节。与其他方法相比较,所提出的盲图像复原方法,从视觉特性和评价指标来讲都具有较好的恢复效果。

When the vehicular image has noise, the estimated blur kernel is not accurate. Therefore a more accurate method based on Variational Dirichlet distribution is proposed to estimate blur kernel, combined with improved augmented Lagrangian to achieve effective image restoration. This method uses the gradient projection method to solve optimization problems and extract precise orientation of the image edge. The Dirichlet distribution substitutes posterior estimate to eliminate image noise and reduce the additional constraint. Hyper - Laplacian prior distribution modeling, together with ALM, is used to restore the vehicular blind image. Experiment results show that multi - scale blur kernel estimator can effectively estimate a blur kernel and eliminate noise of vehicular image, and the texture detail of vehicular image can also be recovered. Compared with other methods, the proposed blind image restoration method have better visual appearances and quality measurements.