基于多尺度MRF图像样本修补稳健算法

摘要

To eleminate cumulative error during image inpainting in single scale, image inpainting robust algorithm based on multiscale MRF model exemplar is proposed. In terms of scale characteristics in image, the sequence maxim posterior probability criterion is applied to inpaint image and corresponding weight cost functions is used for matched sample in the different scale image. Image inpainting information in large scale is adopted to direct the inpainting process in the next more small scale and global and local image inpainting information is combined harmoniously. The experimental results are given, which show that the mosaicing effect and "garbage" which are propagated by cumulative error is suppressed more effectively and the more favourable textural and structural information is preserved.%为了消除在单尺度条件下进行图像修补过程中,在修复区域产生的累积误差,提出了基于多尺度MRF图像样本修补稳健算法.根据图像包含的尺度特征,应用序贯最大后验概率准则对不同尺度的匹配样本采用不同权重的代价函数,利用大尺度填充信息指导小尺度目标区域的修补,实现图像全局信息与局部信息的有机融合.实验结果表明,多尺度MRF图像样本修补算法能更好抑制修复区域由于累积误差产生的“垃圾物”和马赛克现象,同时保持良好的纹理和结构特征.