| 摘要: |
| 目前的图像超分辨技术都依赖于从适当的外部数据集合中提取信息以对图像进行增强,然而这个条件在很多实际应用中难以得到满足.通过对理想边缘模型与纹理内容的分析,发现图像在尺度空间上具有局部结构的自相似性及可传递性.基于这个特点,应用图像类推技术(image analogies,简称IA),可以将图像的局部特性在不同尺度上进行传递,从而为低分辨图像补充结构信息.在实现上,利用原图像和退化图像建立训练集合,用能量图构建学习网络,将图像类推问题转化为求解最小图能量问题.实验结果表明,这种自我类推方法不仅可以有效地提高放大图像的清晰程度,而且较一般的IA算法速度大为加快,更为重要的是,它可以摆脱一般方法对训练集合的依赖,完全独立进行. |
| 关键词: 超分辨 图像类推 马尔可夫网络 |
| DOI： |
| 分类号: |
| 基金项目:Supported by the National Natural Science Foundation of China under Grant No.60473105 (国家自然科学基金); the National Basic Research Program of China under Grant No.2002CB312102 (国家重点基础研究发展计划(973)). |
|
| Image-Analogies Based Super Resolution |
|
GU Yuan-Ting,WU En-Hua
|
| Abstract: |
| Most methods of super resolution so far enhance images by adding exterior information extracted from a given training set. However, this is impractical in lots of cases. From analysis of an ideal edge model and texture contents within images, it is found that many images hold similar local structure at different resolution and preserve it stably in the scale space. Based on this property, Image Analogies can be applied to pass local information onto lower resolution image and thus to achieve resolution enhancement. Original image and its lower-resolution version are used to construct the training set to fit this problem to Image Analogies, and it is resolved by minimizing a graph with energy. Experimental results show that this self analogies algorithm can amplify images much more sharply than traditional interpolation-like methods, and more importantly, it can be executed independently without any supposed outliers. |
| Key words: super resolution image analogies Markov network |
