基于视觉显著性信息的自然图像抠图算法

A natural image matting method based on visual saliency information

为了解决目前抠图算法的采样不充分问题,提出了一种在视觉显著信息指导下的采样,并通过可信度计算确定样本可用性的自然图像抠图算法SIGM.该算法将视觉显著模型引入采样过程,以显著图为指导重点采集区域内的颜色显著样本点.综合分析候选样本对的可信度,将是否符合线性假设、颜色空间内的距离、局部显著性等作为评判依据,选择具有高可信度的样本对来估计透明度值.使用抠图拉普拉斯矩阵作为代价函数的平滑项,优化估计透明度值,得到最终的图像掩膜.实验结果验证了该算法的有效性和准确性.

To solve the problem of insufficient sampling in matting algorithm, we propose a color sam- piing method guided by visual saliency information and matting method based on sample pair confi- dence（SIGM）. This method introduces visual saliency model into the sampling process and collect color salient points. We calculate the confidence of each sample pair using linear hypothesis, color space dis- tance and local saliency as the basis for evaluation. Then we estimate the alpha value with high confi- dence sample pairs. Finally we use the matting Laplace matrix as the smooth item of the cost function to optimize the estimation alpha value and get the final matte. The experimental results show the ef- fectiveness and accuracy of the proposed method.