基于语境相关图传播的图像标注改善方法

提出一种图像标注改善方法,利用数据集蕴含的语境相关信息进行标注改善。构建标签相关图和视觉内容相关图,利用正则化框架将标注改善问题描述为两个无向加权图上的损失函数最小化问题。采用数据分割,逐次优化和放松约束的策略,获得该问题的近似解。该方法充分利用标签的语境相关信息和图像内容相关信息,对数据集分割的粒度具有较好的鲁棒性,具备近似线性的时间复杂度。测试结果表明,该方法适用于大规模数据集,性能优于其它对比方法,可以较大幅度的提升图像标注性能。

Image annotation refinement based on contextual graph diffusion

A new image annotation refinement method was proposed. The initial labels were firstly obtained by annotation methods. Then label relevant graph and visual relevant graph were constructed and mutually reinforced. The semantic optimization problem was formulated into a regularized framework on above undirected weighted graphs. With strategies like data partitioning, successive optimization and constraint relaxation, an approximate optimized solution was got. The refined result could be more related to the content of images by incorporating both visual content and contextual information. Moreover, the proposed method was robust with the partition granularity, and the complexity was approximately linear. Experimental results on large scale web image dataset showed that the proposed method outperformed others, and could achieve both efficiency and capability.