Journal of Frontiers of Computer Science and Technology ›› 2019, Vol. 13 ›› Issue (5): 834-845.DOI: 10.3778/j.issn.1673-9418.1806027

Previous Articles     Next Articles

Salient Object Detection Algorithm Based on Multi-Feature Fusion

ZHANG Shoudong, YANG Ming+, HU Tai   

  1. College of Computer Science and Technology, Nanjing Normal University, Nanjing 210023, China
  • Online:2019-05-01 Published:2019-05-08

基于多特征融合的显著性目标检测算法

张守东,杨  明+,胡  太   

  1. 南京师范大学 计算机科学与技术学院,南京 210023

Abstract: Salient object detection aims at extracting visually salient objects in the image. It is an important task in computer vision and related researches. Considering that many existing algorithms based on deep learning suffer from insufficient feature learning and high detection error rate in complex natural scenes, this paper proposes a novel saliency object detection algorithm based on multi-feature fusion. The features of the saliency map predicted by HDHF (hybrid deep and handcrafted feature) model are used to fuse the deep features of the global pixel. In addition, the candidate nomination is applied to extract the position of candidate objects, and center priors are added to each candidate object. In the fully convolutional neural network, a forward propagation algorithm is utilized to finally predict the pixel-level salient object. Verification is performed on four image datasets with multiple salient objects and complex backgrounds. Experimental results demonstrate that the algorithm effectively improves the detection accuracy of salient objects in complex scenes, especially for the images with complex backgrounds.

Key words: salient object detection, deep learning, complex scene, fully convolutional neural network, multi-feature fusion

摘要: 显著性目标检测是获取图像中视觉显著目标的任务,它是计算机视觉及相关研究领域的重要内容。当前在复杂的自然场景下基于深度学习的算法依然存在特征学习不足和检测错误率较高的问题,因此提出一种新颖的基于多特征融合的显著性目标检测算法。以HDHF(hybrid deep and handcrafted feature)模型的预测显著图作为特征,融合全局像素的深度特征。此外,利用显著性提名获取候选目标的位置,并在各候选目标中添加中心先验。在全卷积神经网络中,利用前向传播算法最终预测得到像素级的显著性目标。在四个包含多个显著性目标和复杂背景的图像数据集上进行验证,实验结果表明,该算法有效地提高了复杂场景下显著性目标的检测精度,尤其是在背景复杂的图像上具有较优的检测效果。

关键词: 显著性目标检测, 深度学习, 复杂场景, 全卷积神经网络, 多特征融合