结合相似性拟合与空间约束的图像分割

目的 图像中的目标一般含有很多子类,仅仅利用某个子类的特征无法完整地分割出目标区域。针对这一问题,提出一种结合相似性拟合与空间约束的图像交互式分割方法。方法 首先,通过手工标记的样本组成各个目标的字典,通过相似度量搜寻测试样本与各个目标的字典中最相似的原子建立拟合项;再结合图像的空间约束项,构建图像分割模型;最后利用连续最大流算法求解,快速实现图像分割的目的。结果 通过对比实验,本文方法的速度比基于稀疏表示的分类方法的速度提高约13倍,而与归一化切割(N-Cut),逻辑回归(logistic regression)等方法相比,本文方法能取得更稳定和准确的分割结果。此外,本文方法无需过完备字典,只需要训练样本能体现各个子类的信息即可得到稳定的图像分割结果。结论 本文交互式图像分割方法,通过结合相似性拟合以及空间约束建立分割模型,并由连续最大流算法求解,实现图像的快速准确的分割。实验结果表明,该方法能够胜任较准确地对自然图像进行分割以及目标提取等任务。

Image segmentation by integrating similarity fitting and spatial constraint

Objective An image cannot be segmented well using one feature in one subclass because the objects in an image region contain different subclasses. The sparse representation classifier (SRC), a combined result of machine learning and compressed sensing, can solve such problems by constructing a dictionary that contains all the features of all subclasses. However, the performance of the SRC depends on an over-complete dictionary. Thus, the SRC needs a large number of training samples to obtain a good segmentation result. Therefore, we propose an interactive segmentation method based on similarity search and spatial constrain. Method First, each object dictionary that contains different features from corresponding subclasses is designed separately by manual labeling. The whole dictionary is then constructed by arranging the object dictionaries obtained by manual labeling in sequence. The most similar feature of the test sample is searched in each dictionary. The fitting term is built using the distance between the testing sample and the most similar feature in each subclass. Afterward, the object function is built by integrating the fitting term and the spatial constraint using total variation. With the addition of the spatial constraint, our model can reduce the effect of noise on the image segmentation results. The continuous max-flow algorithm is applied to minimize the objective function and to effectively obtain segmentation results. The continuous max-flow algorithm, which is linked to the continuous min-cut problem, is a novel method for solving the standard TV-based formulations of the Potts model. This algorithm can avoid extra computational load in enforcing simplex constraints and naturally allows parallel computations over different labels. Result Our method can segment different nature images with different object shapes or contents and is robust to the location and number of training samples. Compared with that of the SRC method, the efficiency of our method is relatively better when using the same dictionary. In addition, the segmentation results of our method are much better than those of traditional segmentation methods, such as N-cut and the logistic regression classifier. These results demonstrate that our method can segment a whole region of different objects in an image. Conclusion The SRC can segment an image that contains different subclasses using an over-complete dictionary and a reconstruction strategy. However, the segmentation performance of the SRC is weak when only a few training samples are used. In this work, the similar feature fitting method and a spatial constraint are used to build a Potts model. The continuous max-flow is applied to solve the objective function and to obtain segmentation results. Our method offers the following advantages. 1) The feature fitting strategy is suitable for image segmentation with a small number of training samples, and it can solve the problem where image objects contain different subclasses. 2) The spatial constraint based on total variation can avoid the noise effect during image segmentation and improve the accuracy of the segmentation results. 3) The continuous max-flow applied to solve our objective function can avoid extra computational load in enforcing simplex constraints and naturally allows parallel computations over different labels. Comparative experiments with SRC, N-cut, and LRC methods demonstrate that our method can segment a whole region with different objects in an image.