基于局部自相似性的视频图像超分辨率算法

Local self-examples based video images super-resolution algorithm

针对视频图像的特点,提出基于局部自相似性的视频图像超分辨率算法。该算法不依赖自然图像数据库作为样本块的图像来源,而是利用局部自相似性,通过在相关坐标邻域中搜索子图像块以实现高频补偿。设计上采样和下采样滤波器,以实现对高频补偿后的图像进行滤波从而产生最终的样本块,采用逐级放大、分多步组合达到视频图像的放大,从而实现了视频图像超分辨率算法。实验结果表明,对于视频序列图像,在主观视觉效果和均方根误差(root mean square error,RMSE)、结构自相似性算子(structural similarity index measurement,SSIM)等方面,算法能显著地提高其分辨率,取得很好的效果。同时,对视频图像利用局部自相似性方法,减少了图像块的检索时间,降低了算法运算量。

According to the features of video images,local self-similarity based video images super resolution algorithm is proposed in the paper. The algorithms do not rely on the external example database,but take the character of local self-similarity to accomplish the high-frequency compensation by searching the example patches in the relative coordinate neighborhood. The up-sample and down-sample filters are designed to filter the compensated images to produce the final example patches,and the progressive magnification is taken to enlarge the video images step by step. Experiments show the algorithm can efficiently improve the resolution of the video images no matter in visual effects or in the quality measures,such as RMSE,SSIM. Meanwhile,the approach of the self-similarity on the video images has reduced the searching time on the image blocks and decreased the algorithm computation.