摘要
研究图像恢复高分辨率图像的问题。由于图像存在锐化和冗余信息而影响了矢量效果,针对以往用反卷积或者单幅图像进行图像恢复重构方法结果不够理想的缺陷,提出了一种利用多幅图像进行图像重构的方法。算法核心思想是利用多幅相关图像进行训练,从而得到某些观测的分布特性,从而对图像进行恢复。首先,定义了一个基于图像小块的马尔科夫随机场分布模型,并由此定义了隐含节点到观测之间的势能函数和隐含节点之间的势能函数;然后通过选用一组训练图像,计算隐含节点到观测之间的混合高斯分布模型,并利用相邻节点所对应的小块之间相邻边界的相关度来计算势能函数所对应的值。最后利用置信传播的方法计算整幅图像上最优的恢复结果。仿真结果显示提出的方法较传统的反卷积方法和插值方法具有更优的恢复结果,能很好的逼近原始的真实图像。
In this paper, we presented a novel method for image restoration where multiple images were used to restore the high resolution image. Firstly, we defined a Markov random field on the low resolution image where two t potential functions were also defined. In the MRF, each node corresponded to a patch in the low resolution image. The first potential function represented the compatibility between hidden node and observations and the second potential function shown the relations between the neighboring hidden nodes. We trained the MRF using a bunch of high resolution images. The Gaussian mixture model over hidden node and observation was estimated based on training images and the similarities between neighboring nodes were calculated according to the edges shared by neighboring nodes. The experiments show that our method can provide better performance of restoration than reverse convolution method and interpolation method.
出处
《计算机仿真》
CSCD
北大核心
2012年第2期285-287,391,共4页
Computer Simulation
