电子学报 ›› 2018, Vol. 46 ›› Issue (11): 2688-2695.DOI: 10.3969/j.issn.0372-2112.2018.11.016

• 学术论文 • 上一篇    下一篇

基于“轮廓-区域”多层互补特性的显著性检测

杨兴明, 王雨廷, 谢昭, 吴克伟   

  1. 合肥工业大学计算机与信息学院, 安徽合肥 230009
  • 收稿日期:2017-09-28 修回日期:2018-02-02 出版日期:2018-11-25
    • 作者简介:
    • 杨兴明 男,1977年5月出生,副教授,合肥工业大学计算机与信息学院,主要研究方向为计算机控制、图像处理、模式识别.E-mail:xmyang168@163.com;王雨廷 女,1993年9月出生,硕士研究生,合肥工业大学计算机与信息学院,主要研究方向为图像处理、模式识别.E-mail:15256952170@163.com;吴克伟 男,1984年9月出生,副研究员,合肥工业大学计算机与信息学院,主要研究方向为计算机视觉、人工智能、模式识别.E-mail:wu_kewei1984@163.com
    • 基金资助:
    • 国家自然科学基金 (No.61273237,No.61503111)

Saliency Detection with Multi-layer Contour-Region Complementary

YANG Xing-ming, WANG Yu-ting, XIE Zhao, WU Ke-wei   

  1. Hefei University of Technology, Hefei, Anhui 230009, China
  • Received:2017-09-28 Revised:2018-02-02 Online:2018-11-25 Published:2018-11-25

摘要: 针对显著性检测在混杂场景中目标容易混淆的问题,本文借助Gestalt心理学理论,利用轮廓线索与外观线索的互补特性,提出一种基于"轮廓-区域"多层互补特性的显著性检测方法.首先,在图像超像素分割基础上,分别提取基于颜色直方图的全局外观线索和基于区域近邻关系的局部对比度线索,充分描述了区域内容的显著性特征;其次,针对混杂场景的区域外观差异小而引起的目标混淆问题,提取基于边缘的目标轮廓封闭性,描述区域轮廓的显著性特征;最后,为了提高对目标尺寸的自适应能力,本文方法使用支持向量机优化多尺度模型中的"轮廓-区域"互补特性融合过程.在ASD,MSRA10K,SED2公认数据集上的实验表明,本文基于轮廓封闭特性的显著性特性,能够有效改善目标显著性查全率、查准率,优于现有的其他先进方法.

关键词: 显著性检测, 轮廓封闭性, 多尺度融合, 外观显著性, 互补性

Abstract: Inspired by Gestalt theory in psychology, we proposes a saliency detection method with multi-layer contour-region characteristic by using the complementary between contour cue and appearance cue, to address the problem that object is indiscoverable in the clutter scene, Firstly, after the image super-pixel segmentation, two kinds of appearance features are extracted respectively, including global appearance cue based on color histogram and local contrast cues based on the regional neighborhood relations. Secondly, in order to solve the object confusion caused by small difference in the appearance, we extract object closure feature to describe the saliency from object contour. Finally, we introduce multi-scale segmentation into our model to enhance the robustness for the diversity of object size, and use the support vector machine to optimize the fusion weights of contour-region cues. Experiments on ASD, MSRA10K, SED2 datasets show that, compared to other state-of-the-art methods, our model can improve the recall and precision measure due to the introduction of contour closure characteristic.

Key words: saliency detection, contour closure, multi-layer fusion, appearance saliency, complementary

中图分类号: