基于海量多源素材优选的多选择图像修复研究

图像修复技术可以对图像缺损区域进行修复，并使得修复结果保持良好的视觉观赏性。尤其，多选择性图像修复利用多种语义素材增加了修复的多选择性、创意性，显著地提升了图像修复技术的实际应用价值。然而，高质量素材图像获取准度较低、素材区域推荐及提取精度较低、修复结果人为痕迹显著等问题降低了多选择性修复的质量。针对上述问题，本申请项目提出了基于海量多源素材优选的多选择图像修复方法。该方法主要包含素材图像精准检索及生成、素材区域精准选取及精度提取、多尺度像素级深度修复等核心模块。多个模块深度结合并相互优化，可以高精度地检索并生成高质量的素材图像，精准推荐及精度提取素材图像中的素材区域，最大限度地减少修复过程中的人为痕迹。最后，在此基础上建立一个基于海量多源素材优选的多选择图像修复平台，在海量图像库中验证理论算法，促进成果应用于数字媒体产业，并在生命科学、教育文化等领域推广。

Image completion technique could complete the missing region, which makes complete result maintain good visual condition. In particular, multi-option completion offers more completion options and creativity to uncompleted image, significantly improves the practical application value of image completion. However, ‘material image obtention is not accurate enough’,‘recommendation and extraction for material region is not precise enough’ and ‘obvious artificial details exist after completion’ problems reduce the quality of completion. Aiming at the problems above, we proposed one massive multi-source material optimized retrieval based multi-option image completion method. The method mainly concludes ‘accurate material image retrieval and generation’, ‘precise material region selection and extraction’, ‘pixel-level multi-completion’ essential modules. They are combined and optimized mutually, which could retrieve and generate high-quality material images accurately, recommend and extract the material regions of material images precisely, minimize the artificial details. At last, on the basis of that, we build up an application platform of massive multi-source material optimized retrieval based multi-option image completion, to verify theoretical algorithms using large database. More importantly, the platform could promote the application of achievements in the digital media industry, apply achievements to fields of life science, education and culture.