Abstract:

Existing image inpainting algorithms based on sparse representation do not distinguish structure patch and texture patch when constructing the learning dictionary.This may increase the processing time when inpainting the smooth part of the damaged region,and decrease the inpainting accuracy when inpainting the structure part.To solve this problem,this paper proposes an improved algorithm.It classifies the patches around the damaged region into two categories where one category includes structure patches while the other includes texture patches,and correspondingly constructs two kind of learning dictionaries.For structure patches,their corresponding dictionaries are used to ensure high accuracy.For texture patches,their corresponding dictionaries are used to accelerate inpainting under the condition of clear inpainted texture.Experimental result shows that the proposed algorithm can effectively inpaint the structure detail without increasing the complexity,while the inpainting effect is better with low time consumption.

Key words: image inpainting, sparse representation, structure patch, texture patch, dictionary construction, structure detail

摘要：

现有基于稀疏表示的图像修复算法在修复破损区域时对纹理块和结构块不加区分,导致修复平滑部分时处理时间较长,同时在修复结构部分时精度较低。针对上述问题,提出一种改进算法。将待修复块分成纹理块和结构块2类,相应构造不同的学习字典。结构块所应用的字典用于确保修复精度,而纹理块所应用的字典则用于在保证纹理清晰的基础上加快修复速度。实验结果表明,该算法在不增加时间复杂度的前提下可有效恢复结构细节,改善破损图像整体修复效果,同时降低处理时间。

关键词: 图像修复, 稀疏表示, 结构块, 纹理块, 字典构造, 结构细节

CLC Number: 

• TP391