样本驱动的半自动图像集前背景分割

Example-Driven Semi-automatic Image Collection Segmentation

图像集的前背景分割是近年来图像处理与图形学领域的一项热点研究工作.针对图像集中的图像逐个进行交互分割会涉及大量的用户操作,导致效率低下,而联合分割方法通常局限于处理具有相似前景的图像集,且因需求解大规模的优化问题较为耗时的问题,提出一种样本驱动的半自动图像集分割方法.首先选取若干图像作为样本进行手动交互分割,训练基于样本图像超像素特征描述的支持向量机分类器;对于其余待分割图像,根据其超像素特征描述到支持向量机分隔超平面的距离计算基于双弯曲Sigmoid函数映射的前景置信度,再采用图切割的算法实现目标图像的快速自动分割.对于包含错误分割的个别图像,进一步提出一种交互式局部修正方法修复错误分割区域,并获得最终的精确分割结果.在2个标准数据集上进行算法有效性验证和对比实验的结果表明,与联合分割算法相比,文中方法能更好、更快地实现在线分割;与逐个交互分割算法相比,文中方法能以相对较小的交互量实现对目标图像集的精确分割.

Binary segmentation for image collection has received considerable attention in image processing and graphic communities recently. Interactively separating foregrounds from an image set one by one is time-consuming and requires tedious user guidance. Meanwhile, image co segmentation techniques generally lack efficiency due to the complexity of solving large optimization problems and are only applicable to images sharing similar foreground appearance. In this paper, we propose an example-driven semi-automatic framework to tackle the image collection segmentation problem. First, we select few sample images from the given image collection and deliver them to user for hand segmentation. Then, super-pixel features based support vector machine （SVM） classifier is trained. For each super-pixel of a given image, we estimate its foreground labeling confidence by applying Sigmoid function on the distance between its descriptor and SVM separation hyperplane. The confidence values are then encoded in a graph cut segmentation procedure to achieve automatic object cutout. For each image with incorrectly segmented regions, accurate result is further obtained by a new proposed local refinement process. Experiments on 2 standard datasets are presented, showingthat the proposed algorithm not only greatly outperforms existing co-segmentation techniques, but also largely reduces users＇ efforts for cutting out object interactively.