改进的规范化卷积图像修复算法  

An Improved Image Inpainting Algorithm Based on Normalized Convolution

作　　者：秦绪佳[1] 桑贤生[1] 程时伟[1] 徐晓刚[2,3] 

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310032 [2]大连舰艇学院装备系统与自动化系,大连116018 [3]浙江大学CAD&CG国家重点实验室,杭州310058

出　　处：《计算机辅助设计与图形学学报》2011年第2期371-376,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61075118;60975016;61005056);浙江省自然科学基金(Y1080436;Y1100880);浙江省科技计划(2009C31106);浙江大学CAD&CG国家重点实验室开放课题(A0906)

摘　　要：规范化卷积算法可用来处理离散、孤立丢失数据的修复,但针对大量数据连续缺失的图像修复效果不佳,为此设计了像素的邻域滤波器,提出一种改进的基于规范化卷积的图像修复算法.首先设定由待修复区域边界向区域内部逐步扩散的修复顺序;然后为每个待修复像素关联一个置信度,并根据修复的先后顺序从高到低进行赋值,即最后修复的像素置信度取值最小;最后根据像素值和置信度进行插值,修复缺失的图像信息.实验结果表明,该算法是一种实用的图像修复算法,可实现图像连续块状区域信息丢失的修复,并可得到良好的修复效果.Although normalized convolution （NC） algorithm can be used to repair discrete isolated data loss, the image inpainting performance for the continuous block data loss using NC isn＇t good. A filter of neighborhood of pixel is defined, and a new image inpainting algorithm based on NC is proposed. The repair order for the loss block data is from the block borders gradually spread to the center. Then, each pixel of the image block is assigned to a confidence, and each pixel is set a confidence value from high to low according to the pixel＇s repairing order. That is the last repaired pixel is set the lowest confidence value. Finally, the pixels to be repaired are interpolated according to the around pixels＇ value and confidence. Experimental results show that the proposed algorithm is a practical image inpainting algorithm, and good performance can be obtained for the continuous data loss image innainting.

关 键 词：规范化卷积 图像修复 插值 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]