基于奖励机制和互斥条件的图像对象协同定位  

Reward strategy based image object co-localization with mutex constraint

作　　者：李劲松[1] 甘朝晖[1] 蒋旻[2] 

机构地区：[1]武汉科技大学信息科学与工程学院,武汉430081 [2]武汉科技大学计算机科学与技术学院,武汉430081

出　　处：《计算机应用研究》2017年第12期3861-3865,3907,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(41571396);国家大学生创新创业训练基金资助项目(201410488017)

摘　　要：随着网络共享图像数据的日益丰富,同类对象检索和定位的需求变得更为迫切,为了更好地运用协同的思想定位同类对象,提出了一种基于奖励机制和互斥条件的对象协同定位方法。该方法采用对象性检测算法的输出结果作为候选,以图结构的描述方式建立模型,将对象在当前图片中的显著性和与其他图片中候选对象的相似性作为奖励,以多个候选在同一图片中的重叠度作为互斥条件,能够无监督地定位出同类对象在各图中的位置。通过广泛的实验,在object discovery数据集和MSRC数据集上评估了该方法并与现有方法进行了对比,取得了较好的定位效果。With the increasingly abundant of images shared on the Internet,the localizing and retrieving of generic object become growing demand. In order to apply the Joint philosophy to localization,this paper proposed an reward strategy based object co-localization method with mutex constraint. Co-localization handled the problem of respectively localizing the common class of object in a set of images. To unsupervised locate the objects which had intra-class variation and inter-class diversity,it formulated a sub set of candidates generated by objectness in a graph structure which treated saliency from background and similarity between each other as reward,together with mutex constraints that could prevent selecting the overlapping region. It performed an evaluation of this method compared to previous approaches on the object discovery datasets and MSRC data set,and got excellent result.

关 键 词：协同定位 奖励机制 互斥条件 图结构 无监督 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]