基于多尺度分解的k邻域随机查找快速图像修复

Multi-scale Decomposition Based k-nearest-neighbor Random Search for Fast Image Completion

该文提出一种基于多尺度分解的k邻域随机查找快速图像修复方法。基于双边滤波下采样分解图像,从图像最粗糙层开始,对每一粗糙层采用基于最小堆的k邻域随机查找算法快速搜索最佳匹配块,利用鲁棒优先级函数确定下一待修复块。每一粗糙层修复后用双边滤波上采样重建下一粗糙层,迭代得到最终的修复结果。与相关工作比较,所提方法的修复结果能够保持图像的细节和边缘信息,取得更高的修复质量。利用客观指标评价修复结果。实验结果表明该方法有效易行,修复的图像具有良好的可视效果。

Multi-scale decomposition based k-nearest-neighbor random search for fast image completion is presented. The image is decomposed using the bilateral filtering based down sampling. Starting from the coarsest level image, the most matching patch is searched using k-nearest-neighbor search algorithm based on the minimum heap for each coarse layer. The robust priority function is presented to determine the next patch that should be handled. The lower coarse layer is reconstructed using the bilateral filtering based up sampling after current coarse layer is repaired, so as to get the final result with iterative completion. Compared with related work, the presented algorithm preserves image details and edge information, and obtains higher completion quality. The completion results are evaluated utilizing the objective indictors. The experimental results show that presented method is effective, feasible, and the visual effect of the image completion is pleasing.