适用于迭代型去模糊算法粗精检测相结合的自适应终止机制

Adaptive coarse and fine combined termination mechanism for iterative deblurring algorithm

由于缺乏合适的去模糊图像质量度量方法,迭代型去模糊算法通常将其迭代次数简单地设置为固定值,无法在执行效率和去模糊质量之间获得最优的平衡点。为此,提出一种粗精检测相结合的迭代终止机制并将其应用到迭代型去模糊算法中以自适应地确定它们最佳的迭代次数。具体地,在每一步迭代过程中利用伪PSNR值细粒度地判断去模糊图像质量是否趋于稳定,另一方面每隔若干步利用从反向卷积残差图像中提取的统计特征值粗粒度准确地判定图像质量是否达到最佳,将两种去模糊图像质量度量方法有机结合以实现一种效率高且准确的迭代终止判定机制。评估结果显示,将所提出的粗精相结合的终止机制应用于NCSR、GSR和ADMM共三种主流去模糊算法后,执行效率可提高50%左右,而去模糊图像质量也得到了最佳保证。实验结果表明,提出的检测机制能够有效地解决各种迭代型去模糊算法因采用固定迭代次数而存在的无益迭代和过迭代问题,非常具有普适性。

Due to the lack of effective blur metric,the existing image deblurring algorithms fixedly set the number of iterations for the sake of simplicity,which results in that they cannot achieve a good tradeoff between the execution efficiency and the deblurred image quality.To this end,this paper proposed a coarse and fine combined termination mechanism,and employed it to adaptively determine the best number of iterations for those iterative deblurring algorithms.Specifically,it used the pseudoPSNR value to quickly determine whether the image quality tended to be stable at each iteration step,and then exploited the statistical feature extracted from reverse convolution residual image at every fixed steps to evaluate the deblurred image quality coarsely yet accurately during the iterative process.The combination of the two image quality metrics made the proposed termination mechanism be efficiency and accuracy for deblurred images.The evaluation results show that,NCSR,GSR,and ADMM deblurring algorithms using the proposed termination mechanism demonstrate a significant improvement over the corresponding original algorithms on the computational efficiency by 50% while obtaining acceptable image quality.Experimental results show that the proposed detection mechanism can effectively solve the useless iteration and over-iteration problems existing in various iterative deblurring algorithms because of the fixed number of iterations,so it is certain universality.