计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (3): 32-40.DOI: 10.3778/j.issn.1002-8331.1607-0134

• 热点与综述 • 上一篇    下一篇

基于深度交叉CNN和免交互GrabCut的显著性检测

杜玉龙,李建增,张  岩,范  聪   

  1. 军械工程学院 无人机工程系,石家庄 050003
  • 出版日期:2017-02-01 发布日期:2017-05-11

Saliency detection based on deep cross CNN and non-interaction GrabCut

DU Yulong, LI Jianzeng, ZHANG Yan, FAN Cong   

  1. Unmanned Aerial Vehicle Engineering Department, Ordnance Engineering College, Shijiazhuang 050003, China
  • Online:2017-02-01 Published:2017-05-11

摘要: 针对传统显著性检测算法特征学习不足,显著性区域边界不明确和检测效果鲁棒性较差等问题,提出一种基于深度交叉卷积神经网络和免交互GrabCut的显著性检测算法。该方法首先针对传统CNN模型中神经元和参数规模较大导致训练困难的不足,根据人眼视觉原理,构建深度交叉卷积神经网络模型(DCCNN);然后,采用超像素聚类方法获取图像区域特征,并通过Beltrami滤波突出图像内的边界特征,利用DCCNN对特征进行学习,在联合条件随机场框架下完成特征融合,实现显著性区域粗糙检测;最后,对粗糙检测结果自适应二值化和形态学膨胀,将显著区域的多边形逼近结果作为GrabCut算法的输入,完成显著性区域的精确检测。实验结果表明所提算法能够有效提高显著性检测精度,具有更好的鲁棒性和普适性。

关键词: 显著性检测, 深度交叉卷积神经网络(CNN), 超像素;Beltrami滤波, 条件随机场(CRF), 免交互GrabCut

Abstract: Aiming at the problem that traditional saliency detection methods suffer from insufficient feature learning, unclear boundary and bad robust detection, a saliency detection algorithm based on deep cross CNN and non-interaction GrabCut is proposed. First, focusing on the hard training caused by the huge amount of neurons and parameters, a deep cross CNN model is proposed combined with the principle of human vision. Thereafter, images regions features are obtained by using superpixels and the edges are extracted by the Beltrami filtering. The features are identified by the DCCNN model and fused under CRF to gain a coarse saliency region. Finally, it binaries and dilates the rough detection results adaptively, inputs the polygonal approximation into GrabCut algorithm to achieve precise saliency detection. Experimental results show that the proposed algorithm improves the detection precision greatly, which is more robust and universal.

Key words: aliency detection, deep cross Convolutional Neural Network(CNN), superpixel, Beltrami filter, Conditional Random Field(CRF), non-interaction GrabCut