基于视觉显著性的场景目标识别  

Scene Object Recognition Based on Visual Saliency

作　　者：薛梦霞[1] 彭晖[2] 刘士荣[2] 张波涛[2] 

机构地区：[1]杭州电子科技大学电气自动化研究所,杭州310018 [2]杭州电子科技大学检测仪表与自动化系统集成技术教育部工程研究中心,杭州310018

出　　处：《控制工程》2016年第5期687-692,共6页Control Engineering of China

基　　金：国家自然科学基金(61175093)

摘　　要：场景目标识别是场景理解的重要内容之一,提出一种新颖的基于视觉显著性的场景目标识别方法。该方法受生物视觉认知机理的启发,利用视觉显著机制来凸显场景中的感兴趣区域。首先利用GBVS模型对图像数据进行高效筛选,获得特定感兴趣的显著性区域。然后基于图分割理论(Graph Cuts Theory)的Grab Cut算法被用于提取图像中的显著性目标。最后运用SURF特征描述目标对象,通过SURF特征的学习产生目标物体的Bo VW视觉词包,运用视觉词包的SVM分类匹配实现目标对象的图像特征与语义描述之间的知识映射。麻省理工学院LabelMe图像库的实验测试结果表明所述方法具有较高的识别率。As scene object recognition is one of the important issues in scene understanding, a novel method is proposed for scene object recognition based on visual saliency. Inspired by biological visual mechanism, visual saliency mechanism is used to highlight interesting areas in the scene. Firstly, GBVS model is introduced to efficiently screen image data to obtain the salient region of significant interest. Then, Grab Cut algorithm based on Graph Cuts Theory is used to extract the salient goals from salient regions. Finally, SURF features are applied to describe the target object, and the bags of visual word for targets are generated through the SURF features learning. The knowledge mapping between the image features and semantic description of the target is achieved by the SVM classifier matching for bags of visual word. The approach is tested by the image library of MIT, and the experiment results demonstrate that the proposed approach has a high rate of recognition.

关 键 词：目标识别 视觉显著性 GrabCut算法 BoVW模型 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]