图像反卷积边界效应的快速抑制算法

Fast Boundary Artifact Reduction Algorithm for Image Deconvolution

基于频域的图像反卷积算法假设输入信号被周期延拓,但边界处的不连续会导致结果产生严重的振铃效应.为此,提出一种快速有效的边界效应抑制算法.首先在现有基于图像延拓的算法的基础上,将延拓区域的定义简化为3种类型;然后提出一种具有对称系数的卷积金字塔滤波器组模型.基于这种模型,在某种特殊图像上针对每种类型的区域各训练一组滤波器组系数,并将由此训练得到的滤波器组用于求解其他图像相应类型的延拓区域.实验结果表明,该算法避免了求解大型稀疏线性方程组,在不影响图像反卷积精度的前提下,可将延拓区域的计算速度提高2个数量级以上,有效地抑制各种频域反卷积算法的振铃效应.

Frequency domain image deconvolution algorithms assume the periodicity of the input signals. But the discontinuities appeared along image boundaries can cause severe ringing artifacts in the restored results. This paper presents a fast and effective boundary artifact reduction method. Inspired by previous expansion based algorithms, this paper firstly simplifies the definition of expansion regions into three types, and then proposes an improved convolution pyramid model with symmetric coefficients. Based on this model, it trains the filter coefficients for each type of region on some special images, before it can be applied to compute expansions for arbitrary images. Experimental results show that the proposed method can avoid solving huge sparse linear systems, and bring an acceleration of two orders of magnitude without sacrificing the accuracy of the deconvolution, while reducing the ringing artifact effectively for all Fourier domain deconvolution methods.