融合多重直方图和SVM的交互式图像分割算法

Interactive Image Segmentation Algorithm Combining Multiple Histogram and SVM

交互式图像分割在图像编辑和医学图像分析等多个领域有着重要的应用。然而,许多交互式图像分割算法高度依赖于用户交互信息,无法利用少量信息精确地提取出目标物体。为了解决上述问题,提出了一种融合直方图和支持向量机(SVM)的交互式图像分割(MHSVM)方法。给定少量的用户输入标记,采用SLIC方法将原始图像分割成若干个不规则区域,同时使用颜色直方图和方向梯度直方图作为每个区域的特征向量,并根据合并规则进行区域合并;然后构建训练样本并平衡正负样本数量,最后协同训练SVM分类器,对剩余未标记超像素进行分类。实验结果表明,MHSVM算法能够从复杂的背景中成功地提取出前景物体。对比其他先进的交互式图像分割算法,MHSVM算法受用户标记的影响更小,且在分割精度上具有明显优势。

Interactive image segmentation has important applications in many fields such as image editing,medical image analysis.However,many interactive image segmentation algorithms are highly dependent on user interaction information,and can-not use a small amount of information to accurately extract the target object.To address the above problems,interactive image seg-mentation combining histogram and support vector machine(MHSVM)is proposed.Given a small number of user input markers,the SLIC method is adopted to segment the original image into several irregular regions,and both the color histogram and gradient orientation histogram are applied as the feature vector of each region,then region merging is done according to the merging rule.Then,and the training samples are constructed,the number of positive and negative samples is balanced.Finally,SVM classifier is co-trained to classify the remaining unlabeled superpixels.The experimental results show that MHSVM extracts foreground objects successfully from the background.MHSVM is less affected by the user input markers in compared with the state-of-the-art interac-tive image segmentation methods,which has obvious advantages in segmentation accuracy.