K近邻分类指导的区域迭代图割算法研究

K-NEAREST NEIGHBOR CLASSIFICATION-GUIDED ITERATIVE REGIONAL GRAPH CUTS ALGORITHM

采用原始图割算法从复杂的背景中提取目标对象,经常需要大量用户交互信息并会产生错误的分割结果。针对此问题,提出一种K近邻分类指导的图割区域迭代分割算法。运用均值漂移算法,将原始图像预分割为多个同质区域作为超像素点。根据用户标记的种子区域构建加权子图,使用图割算法对邻近未标记区域分割标记。利用自训练的K近邻分类器对每次局部分割中新标记超像素点的分割标签进行置信度评估,选择高置信度的超像素点作为新的种子区域指导下一次局部分割。在不同实验图像的分割结果表明,该算法具有良好的准确性和鲁棒性。

Using the original graph cuts algorithm to extract object from complex background often needs a number of user interaction information and gives rise to segmentation errors as well.To solve this problem,we proposed a k-nearest neighbor(KNN)classification-guided iterative regional graph cuts algorithm.With the mean shift algorithm,the original image was pre-segmented into multiple homogeneous regions as super-pixels.A weighted sub-graph was constructed according to the seed labeled by user,and graph cuts algorithm was adopted to segment markers in adjacent unmarked regions.We also used self-training KNN classifier to evaluate the confidence of the segmentation labels with new super pixels in each local segmentation.And super-pixels with high confidence were selected as new seed areas to guide the next local segmentation.The segmentation results of different experimental images show that the proposed method has good accuracy and robustness.