结合CRF与ShapeBM形状先验的图像标记

摘要

Conditional random field (CRF) is a powerful model for image labeling,it is particularly well-suited to model local interactions among adjacent regions (e.g.superpixels).However,CRF doesn't consider the global constraint of objects.The overall shape of the object is used as a global constraint,the ShapeBM can be taken advantage of modeling the global shape of object,and then a new labeling model that combined the above two types of models was presented.The combination of CRF and ShapeBM was based on the superpixels,through the pooling technology was wed to establish the corresponding relationship between the CRF superpixel layer and the ShapeBM input layer.It enhanced the effectiveness of the combination of CRF and ShapeBM and improved the accuracy of the labeling.The experiments on the Penn-Fudan Pedestrians dataset and Caltech-UCSD Birds 200 dataset demonstrate that the model is more effective and efficient than others.%条件随机场(CRF)是一种强大的图像标记模型,适合描述图像相邻区域间(例如超像素)的相互作用.然而,CRF没有考虑标记对象的全局约束.对象的整体形状可以作为对象标记的一种全局约束,利用形状玻尔兹曼机(ShapeBM)在建模对象的整体形状方面的优势,提出了一种CRF与ShapeBM相结合的标记模型.标记模型建立在超像素的基础上,并通过pooling技术在CRF的超像素层与ShapeBM的输入层间建立对应关系,增强了CRF与ShapeBM结合的有效性,提高了标记准确率.在Penn-Fudan Pedestrians数据集和Caltech-UCSD Birds 200数据集上的实验结果表明,联合模型明显地改善了标记结果.