电子学报 ›› 2021, Vol. 49 ›› Issue (3): 596-604.DOI: 10.12263/DZXB.20200161

• 学术论文 • 上一篇    下一篇

Focus+Context语义表征的场景图像分割

吴绿1, 张馨月2, 唐茉2, 王梓3, 王永安4   

  1. 1. 武汉理工大学宽带无线通信与传感器网络湖北省重点实验室, 湖北武汉 430070;
    2. 武汉大学资源与环境科学学院, 湖北武汉 430079;
    3. 中国人民解放军 95028 部队, 湖北武汉 430070;
    4. 中国电子科技集团公司第五十四研究所, 河北石家庄 050081
  • 收稿日期:2020-02-20 修回日期:2020-08-11 出版日期:2021-03-25
    • 作者简介:
    • 吴绿 女,1983年1月出生,湖北潜江人.武汉理工大学通信与信息系统专业博士,现任武汉理工大学信息学院讲师.主要研究方向为机器视觉、模式识别.E-mail:wulv@whut.edu.cn;张馨月 女,1994年7月出生,吉林双辽人.武汉大学资源与环境科学学院博士研究生.
    • 基金资助:
    • "十三五"国家重点研发计划 (No.2017YFB0503500)

Focus+Context Semantic Representation in Scene Segmentation

WU Lü1, ZHANG Xin-yue2, TANG Mo2, WANG Zi3, WANG Yong-an4   

  1. 1. Key Laboratory of Broadband Wireless Communication and Sensor Networks, Wuhan University of Technology, Wuhan, Hubei 430070, China;
    2. School of Resource and Environmental Sciences, Wuhan University, Wuhan, Hubei 430079, China;
    3. A Unit of the Chinese People's Liberation Army, Wuhan, Hubei 430070, China;
    4. The 54 th Research Institute of China Electronics Technology Group Corporation, Shijiazhuang, Hebei 050081, China
  • Received:2020-02-20 Revised:2020-08-11 Online:2021-03-25 Published:2021-03-25

摘要: 场景图像分割一直是机器视觉学习中较为复杂的重难点问题.本文在机器视觉注意力机制学习方法的基础上,融合人类对事物个体的认知,提出场景对象的Focus+Context语义表征,将对象类别信息带入图像底层特征学习中,运用概率统计理论,在抽象层上建模局部区域对象,再联合上下文语义信息推理全局与局部区域对象之间的关系,以实现类内焦点对象(Focus)突出的场景语义分割.实验验证,基于Focus+Context的语义表征和建模能够增加对象的识别率,尤其是在小样本环境下,所提出的方法能极大地简化场景的理解.

 

关键词: 场景图像分割, Focus+Context, 语义表征, 主题模型

Abstract: Scene segmentation has always been a key and complicated problem in machine learning.In order to understand the scene and recognize the objects more accurately,this paper adopts human attention mechanism,takes the category semantic information into consideration and merges it into the image feature learning.The Focus+Context semantic representation is proposed,where the context describes the relationship between the focus and different objects in the scene,and the focus shared among the same category are composed of similar clusters.The probabilistic topic model is used to compute the local features as well as their semantic information.The experimental results show that the Focus+Context method increases the recognition rate of the scene objects,and specially,the proposed method,in a local and global understanding way,can simplify the scene recognition greatly under a small sample size.

 

Key words: scene segmentation, Focus+Context, semantic representation, topic model

中图分类号: