Abstract: Exemplar-based image inpainting methods may cause local mosaic effects and visual incoherence, since the interference of image tiny texture and noise often result in invalid priority terms that makes the inpainting order abnormal. Besides, when searching for the best matching patch, the inter structure information of patches are ignored, which leads to non-unique best matching patches. To tackle these aforementioned issues, a new image inpainting model based on structure-texture decomposition and local total variation minimization was proposed. Three improvements were presented and detailed. Firstly, for an given image to be inpainted, the structure image was extracted by using the logarithm total variation minimization model, then the inpainting priority was calculated on this auxiliary image. In this way, a more robust filling mechanism can be achieved, since the isophote direction of the structure image is less noisy than the original image. Secondly, the priority term was redefined as the weighted summation of data term and confidence term to eliminate the product effect and ensure that the data term was always effective. As a result, the image mismatching rate caused by unreasonable inpainting order was reduced. Finally, the problem of choosing the best matching patch was converted into a 0-1 optimization problem aiming to reach a minimal local total variation. Comprehensive comparisons with the state-of-the-art three inpainting methods show that the Peak Signal-to-Noise Ratio (PSNR) of the proposed algorithm is improved by 1.12-3.56 dB, and the Structural Similarity Index Measure (SSIM) is improved by 0.02-0.04. The proposed model can ensure a better selection of pixel candidates to fill in, and achieve a better global coherence of the reconstruction; therefore, the results are more visually appealing and with less block artifacts for inpainting large damaged images.

Key words: exemplar-based image inpainting, structure-texture decomposition, data term, confidence term, priority term, total variation minimization

摘要： 在基于样例的图像修复算法中，由于优先权公式的计算容易受图像局部噪声和细小纹理的干扰，导致修复顺序错乱；而在搜索最优匹配块时，因忽略了图像块内部的结构影响，可能导致误匹配。针对以上问题提出了一种基于图像的结构-纹理分解及局部总变分最小化的图像修复模型。首先，根据对数总变分最小化模型，将待修复图像进行结构-纹理分解，得到图像的结构分量，并利用图像的结构分量来计算待修复点优先权，使优先权的计算排除局部纹理干扰而更具鲁棒性；其次，将优先权的计算改进为数据项和置信项的加权和，避免了乘积效应，确保数据项一直发挥作用，减少因修复顺序不合理造成的错误匹配；最后，根据图像的局部总变分最小化原则，将图像块的最优匹配转换为0-1优化问题，确保图像修复后的局部结构一致性。与3组参考文献的5组对比实验结果表明，峰值信噪比（PSNR）提高了1.12~3.56 dB，结构相似性指数提高了0.02~0.04。所提模型更好地遵循了修复优先性原则，具有更强的保持图像局部结构一致性的能力，改善了修复图像的视觉效果，适用于复杂结构的大面积毁损的图像的修复。

关键词: 基于样例的图像修复, 结构-纹理分解, 数据项, 置信项, 优先权项, 总变分最小化