| 摘 要: 针对传统显著性目标检测方法在边界模糊以及信息完整性不足等方面存在的问题,提出一种由特征多样性增强模块、全局信息引导模块、特征融合模块组成的显著性目标检测网络。首先,提取各特征层对应的不同类型的空间特征;其次,将坐标注意力机制捕获的信息通过全局引导流传递给各个特征层,学习不同特征层之间的语义关系并且缓解稀释效应;最后,通过级联特征多样性增强模块逐步整合多级特征,实现显著性特征图的精细化生成。在5个公开数据集的对比实验结果表明,该方法有着更优的检测性能,其F-measure和平均绝对误差(MAE)可以分别达到0.959和0.030。 |
| 关键词: 显著性目标检测 全局信息引导 多样性特征 特征融合 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(61801281;61971272) |
|
| Multi-scale Salient Object Detection Network of Integrates Global Guidance Information |
|
YANG Jie1, JIA Xiaoyun1, SHAO Fan2, WEI Lianting1
|
(1.School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China;
2.Xi’an Gaoxin No.1 Experimental High School, Xi’an 710021, China)
3222190133@qq.com; jiaxiaoyun@sust.edu.cn; 2395275662@qq.com; 3236603060@qq.com
|
| Abstract: In order to solve the problems of traditional salient object detection methods in terms of boundary ambiguity and insufficient information integrity, a salient object detection network composed of feature diversity enhancement module, global information guidance module and feature fusion module was proposed. Firstly, different types of spatial features corresponding to each feature layer are extracted. Secondly, the information captured by the coordinate attention mechanism is transmitted to each feature layer through the global guidance stream, so as to learn the semantic relationship between different feature layers and mitigate the dilution effect. Finally, the mult-i level features are gradually integrated through the cascade feature diversity enhancement module to realize the refined generation of salient feature maps. Comparative experimental results on five public datasets show that the proposed method has better detection performance, and its F-measure and Mean Absolute Error(MAE) can reach 0.959 and 0.030,respectively. |
| Keywords: salient object detection global information guidance feature diversity feature fusion |
