基于背景属性的目标识别

Object recognition based on background attributes

属性是图像的语义描述,可以表示图像中某些内容的存在与否,它可以是物体的形状、材质、部件、类别以及功能,也可以是场景的类别以及上下文信息等.由于目标类别与所在背景存在相关关系,提出基于背景属性和目标属性相融合的前景目标识别方法,即对每种背景属性和目标属性分别训练支持向量机(SVM)分类器,并将属性在对应分类器上的得分进行串联组成新的特征,并训练得到最终分类器.对a-Pascal数据库中每幅图像,人工标注了10种背景属性,结合已有的目标属性,进行目标识别实验.与传统方法、基于目标属性的分类方法以及其他前景、背景相结合算法的对比实验结果表明,所提算法比其他算法提高大约2%,背景属性有助于提高目标识别率.

Attribute is the semantic description of an image,which denotes the existence or absence of a semantic property of the image,and it can not only be shape,material,part,category or functionality of an object,but also be label or context of a scene. To improve the accuracy of object classification,considering that object categories are related to the background where they belong to,an approach for object recognition based on modeling background attributes and foreground object attributes was proposed. Each attribute of background and object was trained by a support vector machine( SVM) classifier,and the output value of each attribute classifier was concatenated to form a new feature,based on which the final SVM classifier was trained. 10 kinds of background attributes were manually annotated for each image. Compared to the traditional method,method only based on object attributes and other methods considering different concatenating schemes of background and object features,experiments on the a-Pascal dataset show that the proposed method outperforms the others by around 2%,and background attributes can benefit object recognition task.