基于超像素和SVM的交互式联合分割算法研究

Co-segmentation Based on Superpixel and SVM

联合分割是一类针对前景相同或相似的图像集进行处理的图像分割算法。将分割问题视为前背景像素的分类问题,提出了一种基于超像素和机器学习的联合分割算法,其中使用支持向量机来实现超像素的分类。相比于其他联合分割算法,使用词袋(BOF)模型来描述每个超像素,并引入词频-逆向文件频率(Tf-idf)加权算法来优化提取到的特征。为了减少用户交互工作,通过只在一组前景相似的图像中使用一幅种子图像,并在训练分类器时采用样本抽取的方法来解决正负样本数量不平衡的问题。使用iCoSeg联合分割标准图像集来测试该算法,实验结果表明,相比其他联合分割算法,该方法在精确度和灵活性上都更有优势。

Co-segmentation is defined as segmentation of a set of images with same or similar objects. In this paper, a co-segmentation framework based on superpixel is proposed, and classify superpixels with Support Vector Machine （ SVM ）. In contrast to other co-segmentation methods, the Bag of Features （BOF） model is adopted to describe superpixels and add weight values to descriptors using the term frequency-inverse document frequency （Tf-idf） algorithm. To reduce the interaction of users, one seedimage for one set of images is applied, and balance the number of negative and positive examples by sampling the training data of SVM. Experimental result on iCoSeg dataset shows that the proposed method is more flexible and can achieve better segmentation accuracy, compared with other state-of-the-art co-segmentation methods.