基于Curvelet方向特征的样本块图像修复算法

能否保持修复后图像的结构连贯性和邻域一致性决定了修复性能的优劣.为提高现有样本块修复算法性能,本文提出基于Curvelet变换的样本块图像修复算法.首先利用Curvelet变换估计待修复图像的4方向特征.然后利用颜色信息与方向信息共同衡量样本块间的相似度,在此基础上构造颜色-方向结构稀疏度函数.同时根据构造的加权颜色-方向距离寻找合适的多个匹配块,并利用多个匹配块在构造的颜色和方向空间内的邻域一致性约束下稀疏表示目标块,同时根据目标块所处区域特性自适应确定误差容限.实验结果表明提出算法较现有算法可获得更优的修复效果,尤其是在修复富含结构纹理破损类型的图像时.

Exemplar Based Image Inpainting Algorithm Using Direction Features of Curvelet Transform

Whether the structure coherence and neighborhood consistency can be well maintained directly determines the performance of an inpainting algorithm.To achieve a better inpainting performance, this paper proposes an exemplar based image inpainting algorithm based on direction features extracted by Curvelet transform.Firstly, the super-wavelet transform is applied to extract four direction features of the corrupted image.Then the color and direction information are utilized to measure the similarities between patches.Subsequently, a color-direction structure sparsity function is defined.Afterwards, multiple suitable candidate patches are searched based on the weighted color-direction distance and these candidate patches are applied to sparsely represent target patch under the local neighborhood consistence constraints both in color and direction spaces.Moreover, in searching candidate patches, the error tolerance is adaptively decided according to the feature of target patch.Experiment results show that the proposed method can achieve better inpainted results than the state-of-the-art algorithms, especially when dealing with structure and texture images.