CAJ | 학술논문

针对图像合成中源图像与目标图像光照环境不一致造成直接合成图像不逼真的问题，提出一种基于马尔科夫随机场的光照一致图像合成方法。首先基于加权的泊松克隆方法构建梯度保持的光滑约束，削弱传统的泊松克隆方法在合成边界源图像和目标图像光照差异变化剧烈时产生的渗透效应；然后基于直方图对齐的方法构建光照一致的数据约束，保持合成图像前、背景亮度主轴的一致性；最后根据合成边界源图像的边缘特性以及源图像和目标图像光照差异变化的剧烈程度自适应地调整2项约束的权重，并采用融合局部和全局一致性的学习算法对构建的马尔科夫随机场函数进行快速求解。实验结果表明，该方法产生的合成效果在梯度特征保持方面以及亮度一致性方面均优于传统的泊松克隆方法，同时收敛速度得到了提高。
침대도상합성중원도상여목표도상광조배경불일치조성직접합성도상불핍진적문제，제출일충기우마이과부수궤장적광조일치도상합성방법。수선기우가권적박송극륭방법구건제도보지적광활약속，삭약전통적박송극륭방법재합성변계원도상화목표도상광조차이변화극렬시산생적삼투효응；연후기우직방도대제적방법구건광조일치적수거약속，보지합성도상전、배경량도주축적일치성；최후근거합성변계원도상적변연특성이급원도상화목표도상광조차이변화적극렬정도자괄응지조정2항약속적권중，병채용융합국부화전국일치성적학습산법대구건적마이과부수궤장함수진행쾌속구해。실험결과표명，해방법산생적합성효과재제도특정보지방면이급량도일치성방면균우우전통적박송극륭방법，동시수렴속도득도료제고。
Directly merging a source image into a target image with different lighting condition often leads to unnatural composition result. This work proposes a luminance consistent image composition method based on Markov random field. First, the method builds a gradient maintaining contrast term based on weighted Poisson cloning method. Comparing to the traditional Poisson cloning algorithm, the proposed method easily weakens the bleeding effect in case of serious changes in luminance difference between the source and target image among the composition boundary. Second, the method builds a luminance consistent data term based on the histogram alignment method so as to make the source image’s main luminance axis align with the target image. Then, the method adaptively combines two terms based on the contrast feature of the source image and the change range in luminance difference between the source and target image among the composition boundary. Finally, the method uses the Learning with Local and Global Consistency algorithm for a fast solution of the Markov random field problem. Experimental results demonstrate that the proposed method outperforms the traditional Poisson cloning method in terms of maintaining source image’s gradient and luminance consistency, while achieving a faster con-vergence speed.