计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (14): 199-206.DOI: 10.3778/j.issn.1002-8331.1906-0357

• 图形图像处理 • 上一篇    下一篇

结合边缘特征先验引导的深度卷积显著性检测

时斐斐,张松龙,彭力   

  1. 江南大学 物联网工程学院 物联网技术应用教育部工程研究中心,江苏 无锡 214122
  • 出版日期:2020-07-15 发布日期:2020-07-14

Deep Convolution Saliency Detection Combined with Edge Feature Prior-Based Inspection

SHI Feifei, ZHANG Songlong, PENG Li   

  1. Engineering Research Center of Internet of Things Applied Technology, Ministry of Education, College of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2020-07-15 Published:2020-07-14

摘要:

针对当前基于深度学习的显著性检测算法缺少利用先验特征和边缘信息,且在复杂场景中难以检测出鲁棒性强的显著性区域的问题,提出了一种结合边缘特征,利用先验信息引导的全卷积神经网络显著性检测算法。该算法利用三种被经常用到的先验知识结合边缘信息形成先验图,通过注意力机制将提取的先验特征与深度特征有效融合,最终通过提出的循环卷积反馈优化策略迭代地学习改进显著性区域,从而产生更可靠的最终显著图预测。经过实验定性定量分析,对比证明了算法的可靠性。

关键词: 显著性检测, 全卷积网络, 先验信息引导, 循环卷积优化

Abstract:

In view of the lack of using prior features and edge information in the current significance detection algorithm based on deep learning, it is difficult to detect the significant regions with strong robustness in complex scenes, a method for significance detection of total convolutional neural network based on edge features and prior information is proposed. Firstly, the algorithm uses three kinds of frequently used prior knowledge combined with edge information to form a priori map, and the priori map features and depth features are effectively merged through the attention mechanism. Finally, the circular convolution-feedback optimization strategy is proposed to improve the significance area iteratively, so as to produce a more reliable prediction of the final significance graph. In the end, the reliability of the algorithm is proved by experimental qualitative and quantitative analysis.

Key words: saliency detection, full convolutional network, priori information guidance, circular convolution-feedback optimization