计算机应用 ›› 2018, Vol. 38 ›› Issue (1): 91-96.DOI: 10.11772/j.issn.1001-9081.2017061633

• 人工智能 • 上一篇    下一篇

基于多任务深度卷积神经网络的显著性对象检测算法

杨帆, 李建平, 李鑫, 陈雷霆   

  1. 电子科技大学 计算机科学与工程学院, 成都 611731
  • 收稿日期:2017-07-03 修回日期:2017-08-23 出版日期:2018-01-10 发布日期:2018-01-22
  • 通讯作者: 杨帆
  • 作者简介:杨帆(1987-),男,四川简阳人,博士研究生,主要研究方向:计算机视觉、深度学习、语义级稠密匹配;李建平(1964-),男,湖南祁阳人,教授,博士生导师,博士,主要研究方向:小波信号处理、模式识别、图像处理;李鑫(1986-),男,四川江油人,博士研究生,主要研究方向:计算机视觉、深度学习、人工智能;陈雷霆(1966-),男,重庆人,教授,博士生导师,博士,主要研究方向:图形处理、多媒体技术、图像处理。
  • 基金资助:
    国家自然科学基金资助项目(61370073);国家863计划项目(2015AA016010)。

Salient object detection algorithm based on multi-task deep convolutional neural network

YANG Fan, LI Jianping, LI Xin, CHEN Leiting   

  1. School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu Sichuan 611731, China
  • Received:2017-07-03 Revised:2017-08-23 Online:2018-01-10 Published:2018-01-22
  • Supported by:
    This work is partially supported by the National Natural Science Foundation (6157021026), the National High Technology Research and Development Program (863 Program) of China (2015AA016010).

摘要: 针对当前基于深度学习的显著性对象检测算法不能准确保存对象边缘的区域,从而导致检测出的显著性对象边缘区域模糊、准确率不高的问题,提出了一种基于多任务深度学习模型的显著性对象检测算法。首先,基于深度卷积神经网络(CNN),训练一个多任务模型分别学习显著性对象的区域和边缘的特征;然后,利用检测到的边缘生成大量候选区域,再结合显著性区域检测的结果对候选区域进行排序和计算权值;最后提取出完整的显著性图。在三个常用标准数据集上的实验结果表明,所提方法获得了更高的准确率,其中F-measure比基于深度学习的算法平均提高了1.9%,而平均绝对误差(MAE)平均降低了12.6%。

关键词: 显著性对象检测, 深度学习, 边缘检测, 多任务神经网络, 显著图, 卷积神经网络

Abstract: The current deep learning-based salient object detection algorithms fail to produce accurate object boundaries, which makes the regions along object contours blurred and inaccurate. To solve the problem, a salient object detection algorithm based on multi-task deep learning model was proposed. Firstly, based on deep Convolutional Neural Network (CNN), a multi-task model was used to separately learn region and boundary features of a salient object. Secondly, the detected object boundaries were utilized to produce a number of region candidates. After that the region candidates were re-ranked and their weights were computed by combining the results of salient region detection. Finally, the entire saliency map was extracted. The experimental results on three widely-used benchmarks show that the proposed method achieves better accuracy. According to F-measure, the proposed method averagely outperforms the deep learning-based algorithm by 1.9%, while lowers the Mean Absolutely Error (MAE) by 12.6%.

Key words: salient object detection, deep learning, boundary detection, multi-task deep neural network, saliency map, Convolutional Neural Network (CNN)

中图分类号: