\r何 凯，沈成南，刘 坤，高圣楠\r
\r
AuthorsHTML:\r何 凯，沈成南，刘 坤，高圣楠\r
\r
AuthorsListE:\rHe Kai，Shen Chengnan，Liu Kun，Gao Shengnan\r
\r
AuthorsHTMLE:\rHe Kai，Shen Chengnan，Liu Kun，Gao Shengnan\r
\r
Unit:\r天津大学电气自动化与信息工程学院，天津 300072\r
\r
Unit_EngLish:\rSchool of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China\r
\r
Abstract_Chinese:\r基于马尔可夫随机场(MRF)的图像修复算法，在纹理和结构区域均能获得较好的修复效果．然而，基于MRF 对图像进行修复，各节点存在大量近似的候选块．传统基于MRF 修复算法需要对各节点的近似候选块进行多次重复计算，执行效率低、计算量较大．为克服这一缺点，在马尔可夫随机场框架下，提出了一种快速图像修复算法．在初次迭代前，首先对破损图像进行预处理，采用自适应样本块修复算法，对高斯金字塔顶层的低分辨率图像进行快速的“预修复”，以粗略估计破损区域中MRF 内部节点的初始值，加快后续相邻节点间的消息传递及收敛速度．其次，以“预修复”结果中的初始信息为约束条件，提出了改进的置信度计算方法．同时，将初始置信度最高的候选块设为节点的第一候选块，根据预设的相似度判别阈值，并利用破损块源区域的纹理复杂程度，对MRF 节点的候选块进行筛选，以避免同一个节点具有大量相似的候选块，提高节点的交互运算效率．最后利用MRF 进行迭代计算，获得各节点的最优匹配块，实现图像的自动修复．实验仿真结果表明：与传统基于MRF 修复算法相比，改进后算法的平均运算时间减少了75%以上，可以获得更高的峰值信噪比(PSNR)，修复效果也更为理想；在提高修复效率的同时，取得了更理想的修复效果．\r
\r
Abstract_English:\rImaging inpainting algorithms using Markov random field(MRF)perform well in both textural and structural regions．However，numerous approximate candidate patches often exist in each node．Traditional MRF-based inpainting algorithms are inefficient and time-consuming as multiple approximate candidate patches are repeatedly calculated for every node．To overcome this problem，a fast image inpainting algorithm using MRF framework is proposed herein．First，the damaged image was preprocessed prior to the initial iteration，and the adaptive patch inpainting algorithm was used to process the low-resolution image on the top level of the Gaussian pyramid．In this manner，the initial values of MRF internal nodes in the damaged region could be roughly estimated thereby accelerating the subsequent message transmission between neighbor nodes and convergence．Second，an improved belief calculation method was proposed using the initial information obtained from the “pre-inpainting” result. Simultaneously，the candidate patch with the highest initial belief was set as the first one for every node，and the candidate patches of the MRF nodes were selected according to the preset similarity threshold and the texture complexity in the source region of damaged patches．In this manner，multiple similar candidate patches can be avoided in a node thereby improving the interaction efficiency among nodes．Finally，the optimal matched patch of each node was determined using iterative computations of MRF and automatic image inpainting was realized．Experimental results show that compared with traditional algorithms，the improved one has lower time consumption(reduced by more than 75% on average)，higher peak signal-to-noise ratio(PSNR)，and more reasonable inpainting effects．Therefore，better inpainting effects can be efficiently achieved using the improved algorithm．\r
\r
Keyword_Chinese:图像修复；马尔可夫随机场；置信度传播；高斯金字塔\r

Keywords_English:image inpainting；Markov random field(MRF)；belief propagation；Gaussian pyramid\r


PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6389