Journal of Frontiers of Computer Science and Technology ›› 2018, Vol. 12 ›› Issue (9): 1454-1464.DOI: 10.3778/j.issn.1673-9418.1708023

Previous Articles     Next Articles

Cosaliency Detection Based on Similar Matrix and Clustering Consistency

ZHENG Haijun1,2,3, WU Jianguo1,2,3, LIU Zhengyi1,2,3+   

  1. 1. Key Laboratory of Intelligent Computing & Signal Processing, Ministry of Education, Anhui University, Hefei 230601, China
    2. School of Computer Science and Technology, Anhui University, Hefei 230601, China
    3. Co-Innovation Center for Information Supply & Assurance Technology, Anhui University, Hefei 230601, China
  • Online:2018-09-01 Published:2018-09-10

相似矩阵和聚类一致性的协同显著检测

郑海军1,2,3,吴建国1,2,3,刘政怡1,2,3+   

  1. 1. 安徽大学 计算智能与信号处理教育部重点实验室,合肥 230601
    2. 安徽大学 计算机科学与技术学院,合肥 230601
    3. 安徽大学 信息保障技术协同创新中心,合肥 230601

Abstract: In view of the shortcomings of existing cosaliency detection algorithm, an algorithm based on similar matrix decomposition model and clustering consistency is proposed. Firstly, saliency thresholding is performed to extract the saliency regions for each image by selecting salient superpixels based on different maps. Secondly, the histogram vector is constructed by the RGB color feature of the saliency region, and all the row vectors are combined into a feature matrix, then histogram??s similar matrix is constructed. Thirdly, the low rank matrix decomposition model is used to decompose the feature matrix to obtain sparse matrix and weighted value. Weight value and elemental saliency maps are fused to get the weighted saliency maps. Finally, the final cosaliency maps are obtained by combining the clustering consistency maps and weighted saliency maps. The experimental results on iCosegSub and iCoseg image datasets, demonstrate that the method has better precision- recall curve and higher precision.

Key words: cosaliency detection, low rank matrix decomposition, clustering consistency, similar matrix

摘要: 针对现有协同检测算法存在显著非协同目标抑制不足的问题,提出了一种相似矩阵和聚类一致性的协同显著目标检测算法。首先,对图像进行超像素分割和从现有的多种显著检测算法得到基本显著图,通过基本显著图设置阈值结合超像素提取相应显著区域;其次,通过显著区域的RGB颜色特征构造直方图行向量,并将所有行向量组合成一个特征矩阵,再根据所有直方图应具有相似性特征并以此构造相似矩阵;再次,将相似矩阵应用于低秩矩阵分解模型中,分解特征矩阵得到噪音稀疏矩阵,并以此得到加权值,融合加权值和基本显著图得到加权显著图;最后,采用聚类一致性得到协同显著值,并融合加权显著值,得到最终显著图。在iCosegSub和iCoseg数据集验证实验,算法更好地抑制了显著非协同区域,取得了较高的准确度。

关键词: 协同显著检测, 低秩矩阵分解, 聚类一致性, 相似矩阵