Attention-based bi-directional refinement network for salient object detection

In the past few years, with the development of the fully convolutional neural network, salient object detection has been developed rapidly, whereas the detection accuracy in complex scenes becomes a big challenge. Several features optimization techniques have been developed including cross-optimizin...

Full description

Saved in:
Bibliographic Details
Published inApplied intelligence (Dordrecht, Netherlands) Vol. 52; no. 12; pp. 14349 - 14361
Main Authors Yuan, JunBin, Wei, Jinhui, Wattanachote, Kanoksak, Zeng, Kun, Luo, Xiaonan, Xu, Qingzhen, Gong, Yongyi
Format Journal Article
LanguageEnglish
Published New York Springer US 01.09.2022
Springer Nature B.V
Subjects
Artificial Intelligence
Artificial neural networks
Computer Science
Edge detection
Machines
Manufacturing
Mechanical Engineering
Modules
Object recognition
Optimization
Optimization techniques
Processes
Salience
Multi-scale object capture
Salient object detection
Bi-directional refinement
Gaussian attention
Online AccessGet full text

Cover

More Information
Summary:In the past few years, with the development of the fully convolutional neural network, salient object detection has been developed rapidly, whereas the detection accuracy in complex scenes becomes a big challenge. Several features optimization techniques have been developed including cross-optimizing edge features and salient features, to improve the accuracy of detection. However, in challenging scenes, there still exit irregular detection and segmentation problems caused by noise misjudgment. To alleviate this issue, we proposed an attention-based bi-directional refinement network (ABRN) for salient object detection. Our proposed technique implemented a Gaussian attention module (GAM) to preprocess the features which aims to reduce the noise in salient and edge detection, and to selectively attend to the object in detection process. Moreover, the feature bi-directional refinement module (FBRM) was implemented to refine salient and edge features for each other. Furthermore, the multi-scale object capture module (MOCM) was adapted to dilate the receptive field, which reduces the information loss in convolution process. The experimental results show that the proposed model outperforms 13 state-of-the-art methods on five mainstream benchmark datasets.
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-021-03040-8