Journal of Computer Applications ›› 2017, Vol. 37 ›› Issue (10): 2841-2846.DOI: 10.11772/j.issn.1001-9081.2017.10.2841

Previous Articles     Next Articles

Image labeling based on fully-connected conditional random field

LIU Tong, HUANG Xiutian, MA Jianshe, SU Ping   

  1. Graduate School at Shenzhen, Tsinghua University, Shenzhen Guangdong 518055, China
  • Received:2017-04-10 Revised:2017-05-22 Online:2017-10-10 Published:2017-10-16
  • Supported by:
    This work is partially supported by the National High Technology Research and Development Program (863 Program) of China (2015AA043302), the Program of Industry-university-research Cooperation in Guangdong Province (2013A090100002), the Development Program of Technology Research in Guangdong Huadu District (HD14ZD004).


刘彤, 黄修添, 马建设, 苏萍   

  1. 清华大学 深圳研究生院, 广东 深圳 518055
  • 通讯作者: 黄修添(1991-),男,福建宁德人,硕士研究生,主要研究方向:模式识别、计算机视觉,
  • 作者简介:刘彤(1970-),男,河南洛阳人,讲师,博士,主要研究方向:计算机视觉、LED封装;黄修添(1991-),男,福建宁德人,硕士研究生,主要研究方向:模式识别、计算机视觉;马建设(1965-),男,河南郑州人,副教授,博士,主要研究方向:计算机视觉、机器人控制;苏萍(1975-),女,河南洛阳人,讲师,博士,主要研究方向:三维全息显示系统、图像处理.
  • 基金资助:

Abstract: The traditional image labeling models often have two deficiencies; they only can model short-range contextual information in pixel-level of the image and have a complicated inference. To improve the precision of image labeling, the fully-connected Conditional Random Field (CRF) model was used; to simplify the inference of the model, the mean filed approximation based on Gaussian kd-tree for inference was proposed. To verify the effectiveness of the proposed algorithm, the experimental image datasets not only contained the standard picture library MSRC-9, but also contained MyDataset_1 (machine parts) and MyDataset_2 (office table) which made by authors. The precisions of the proposed method on those three datasets are 77.96%, 97.15% and 95.35% respectively, and the mean cost time of each picture is 2s. The results indicate that the fully-connected CRF model can improve the precision of image labeling by considering the contextual information of image and the mean field approximation using Gaussian kd-tree can raise the efficiency of inference.

Key words: Conditional Random Field (CRF), image labeling, contextual information, Gaussian kd-tree, model inference

摘要: 传统的图像标注模型通常存在两个问题:只能够对短距离的像素上下文信息进行建模和复杂的模型推理过程。为了提高图像标注的精度、简化图像标注的模型推理过程,采用完全联系的条件随机场模型进行图像标注,提出利用基于高斯kd树的平均场估计方法实现该模型的高效推理。为了更好地验证算法的有效性,实验的图片数据库不仅包含标准的图片库--剑桥大学微软研究图片库(MSRC-9),还包含作者制作的机械零件图片库(MyDataset_1)和办公桌图片库(MyDataset_2)。新算法在三个图片库上的平均标注精度分别可以达到77.96%、97.15%和95.35%,每幅图的平均运行时间为2s。实验结果表明,基于完全联系的条件随机场的图像标注能够更充分地考虑不同的像素上下文信息来提高标注精度,而基于高斯kd树的模型推理能够提高模型推理的效率。

关键词: 条件随机场, 图像标注, 上下文信息, 高斯kd树, 模型推理

CLC Number: