高级检索
    梁大川, 李静, 刘赛, 李东民. 基于图和稀疏主成分分析的多目标显著性检测[J]. 计算机研究与发展, 2018, 55(5): 1078-1089. DOI: 10.7544/issn1000-1239.2018.20160681
    引用本文: 梁大川, 李静, 刘赛, 李东民. 基于图和稀疏主成分分析的多目标显著性检测[J]. 计算机研究与发展, 2018, 55(5): 1078-1089. DOI: 10.7544/issn1000-1239.2018.20160681
    Liang Dachuan, Li Jing, Liu Sai, Li Dongmin. Multiple Object Saliency Detection Based on Graph and Sparse Principal Component Analysis[J]. Journal of Computer Research and Development, 2018, 55(5): 1078-1089. DOI: 10.7544/issn1000-1239.2018.20160681
    Citation: Liang Dachuan, Li Jing, Liu Sai, Li Dongmin. Multiple Object Saliency Detection Based on Graph and Sparse Principal Component Analysis[J]. Journal of Computer Research and Development, 2018, 55(5): 1078-1089. DOI: 10.7544/issn1000-1239.2018.20160681

    基于图和稀疏主成分分析的多目标显著性检测

    Multiple Object Saliency Detection Based on Graph and Sparse Principal Component Analysis

    • 摘要: 针对具有多个显著目标且背景较为复杂的图像,提出了一种基于全连接图和稀疏主成分分析(sparse principal component analysis, sPCA)的显著性检测方法.首先,在不同的尺度空间上利用目标先验知识快速获取包含预选显著目标的空间位置信息,同时,在超像素分割的基础上构造全连接图,并计算超像素级的显著图.然后,利用目标先验知识提取并优化超像素显著图的显著性区域,采用稀疏主成分分析提取优化后的显著性像素点的主要特征,获取相应尺度的显著图.最后,将多个尺度下的显著图进行融合得到最终的显著图.该方法充分利用了超像素与像素显著性计算的优势,在提高检测速度的同时获得更高的检测精度.在公开的多目标数据集SED2和HKU_IS上进行实验验证,结果表明:该方法能够有效检测出复杂背景下的多个显著目标.

       

      Abstract: In order to detect multiple salient objects from the image with cluttered background, a new multi-object salient detection method based on fully connected graph and sparse principal component analysis is proposed. Firstly, a rapid coarse detection method with different scales is adopted to obtain the object prior with the location of candidate objects and the pixel level saliency map. Meanwhile, we construct a fully connected graph based on the superpixel segmentation to obtain the superpixel-level saliency map. The salient regions are extracted from the superpixel-level binarized salient object prior map and a sparse principal component analysis method is used to gain the main features vector from the pixel matrix composed of the pixels in the optimized salient regions and obtain the salient map of corresponding scale. Finally, the final salient map is fused with the multi-scale saliency maps. Our method takes the advantage of pixel and superpixel method, it can not only simplify the calculation but also improve the detection precision of the salient objects in the image. Quantitative experiments on two public datasets SED2 and HKU_IS demonstrate that out method can detect multiple salient objects from complex images and outperforms other state-of-the-art methods.

       

    /

    返回文章
    返回