融合空间上下文的自然场景语义建模  

Natural scene classification based on spatial context of local semantics

作　　者：程环环[1] 王润生[1] 

机构地区：[1]国防科技大学电子科学与工程学院ATR重点实验室,湖南长沙410073

出　　处：《电路与系统学报》2010年第6期39-46,51,共9页Journal of Circuits and Systems

基　　金：国家自然科学基金(60373000)

摘　　要：针对自然场景图像,本文提出一种融合空间上下文的场景语义建模和分类方法。针对场景中的局部语义对象,建立了基于贝叶斯网络的语义上下文模型。通过对已标注训练样本集的学习训练,获得局部语义对象在各类场景下的上下文模型。对于待分类的图像,首先利用支持向量机实现分割区域的分类,根据学习得到的语义上下文模型,提取图像中各语义对象的空间上下文信息,形成图像的语义上下文描述,实现场景分类。针对不同场景下的局部语义对象,利用贝叶斯网络自动学习得到不同的空间关系集合用于上下文信息提取,使得场景描述和分类过程更智能和有效。通过在六类自然场景图像数据集上的实验表明,本文所提算法能够很好的利用上下文信息,并取得满意的分类结果。In this paper, a novel semantic modeling and classification method based on spatial context is presented for natural scenes. A semantic-context model based on Bayesian network （SCBN） is presented for local semantic objects in scenes. The context model for each local semantic object in scenes is learned from the training set with manual annotations. For test images, segmented regions are classified into object classes by SVM. Then, spatial contexts of each semantic object are extracted by the learned SCBN models. Images are represented through the frequency of occurrence of these semantic objects and their spatial contexts. The spatial context nodes of SCBN model vary with different scenes and local semantic objects, which are learned by the structure learning algorithm for Bayesian networks. Therefore, our scene description is capable of using specific context for each scene type, which makes the classification process more intelligent and eff^cient. Experiment conducted on natural scenes＇ dataset demonstrates the effectiveness and efficiency of the proposed approach for semantic context modeling and categorization of natural scenes.

关 键 词：贝叶斯网络 场景分类 语义建模 空间上下文 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]