基于回归预测集成学习的交互式图像分割（英文）

发布时间：2018-10-31 14:23

【摘要】：对于复杂场景下的自然图像,全自动图像分割方法难以获得与真实情况吻合的结果,人们常常采用交互式分割手段实现精确分割。然而,当前及背景中存在颜色相似的区域时,传统半监督图像分割方法只能通过大量增加手工标记获得精确分割结果。为此,本文提出一种结合半监督学习的基于回归预测的集成学习交互式图像分割方法。通过集成两个互补的样条回归函数,将图像分割视为一个非线性预测问题。首先,基于已标记样本训练出两个在属性上互补的多元自适应回归样条学习器(multivariate adaptive regression splines,MARS)和薄板样条回归学习器(thin plate spline regression,TPSR);接着,提出一种基于聚类假设和半监督学习的回归器增强算法,该算法从未标记样本中抽选部分样本辅助训练MARS和TPSR;然后,引入支持向量回归方法(support vector regression,SVR)集成MARS和TPSR的预测结果;最后,对SVR集成结果进行Graph Cut图像分割。在标准数据库BSDS500和Pascal VOC上进行大量实验,验证了所提算法的有效性。大量对比实验证实,所提算法在交互式自然图像分割上的表现与当前最先进算法相当。
[Abstract]:For natural images in complex scenes, automatic image segmentation methods are difficult to obtain results consistent with the real situation, people often use interactive segmentation methods to achieve accurate segmentation. However, when there are similar color regions in the background and present, the traditional semi-supervised image segmentation method can only obtain accurate segmentation results by adding a large number of manual markers. In this paper, an integrated learning interactive image segmentation method based on regression prediction combined with semi-supervised learning is proposed. Image segmentation is regarded as a nonlinear prediction problem by integrating two complementary spline regression functions. Firstly, based on the labeled samples, two multivariate adaptive regression spline Learner (multivariate adaptive regression splines,MARS) and thin Spline regression Learner (thin plate spline regression,TPSR) are trained. Then, a regression enhancement algorithm based on clustering hypothesis and semi-supervised learning is proposed. The algorithm has never labeled the samples selected from the samples to assist in training MARS and TPSR;. Then, support vector regression (support vector regression,SVR) is introduced to integrate the prediction results of MARS and TPSR. Finally, the Graph Cut image segmentation of SVR ensemble results is carried out. A large number of experiments are carried out on the standard database BSDS500 and Pascal VOC to verify the effectiveness of the proposed algorithm. A large number of comparative experiments show that the proposed algorithm is comparable to the most advanced algorithms in interactive natural image segmentation.
【作者单位】： School
【基金】：supported by the National Natural Science Foundation of China(Nos.61071176,61171192,and 61272337) the Doctoral Fund of the Ministry of Education of China(No.20130162110013)
【分类号】：TP391.41

【相似文献】