电子学报 ›› 2022, Vol. 50 ›› Issue (8): 1894-1904.DOI: 10.12263/DZXB.20211123

• 学术论文 • 上一篇    下一篇

小目标特征增强图像分割算法

任莎莎, 刘琼()   

  1. 华南理工大学校软件学院,广东 广州 511436
  • 收稿日期:2021-08-18 修回日期:2022-06-22 出版日期:2022-08-25
    • 通讯作者:
    • 刘琼
    • 作者简介:
    • 任莎莎 女,1992年3月出生,安徽淮北人.现为华南理工大学软件学院在读博士研究生.主要研究方向为信号处理、图像理解与分割等方向.E-mail: 201910107240@mail.scut.edu.cn
      刘 琼 女,1959年3月出生,云南昆明人.现为华南理工大学软件学院教授、博士生导师.承担、参加国家自然科学基金项目、国家863、973及地方政府项目10余项,国内外学术刊物和会议发表论文50余篇,中国专利5件.研究方向为计算机网络、计算机视觉、模式识别等.
    • 基金资助:
    • 国家自然科学基金(61976094);广东省自然科学基金(2021A1515011349)

A Tiny Target Feature Enhancement Algorithm for Semantic Segmentation

REN Sha-sha, LIU Qiong()   

  1. School of Software Engineering, South China University of Technology, Guangzhou, Guangdong 511436, China
  • Received:2021-08-18 Revised:2022-06-22 Online:2022-08-25 Published:2022-09-08
    • Corresponding author:
    • LIU Qiong
    • Supported by:
    • National Natural Science Foundation of China(61976094);National Natural Science Foundation of Guangdong Province(2021A1515011349)

摘要:

在图像场景分割中存在小目标易丢失,边缘轮廓噪声大等问题.在目前的增强特征表征能力与优化空间细节的语义分割算法中,由于边缘和小目标特征的丢失,导致小目标和边缘很难被准确分割.为此,本文研究了一种小目标特征增强的图像分割算法.首先设计一种像素空间注意力模块(Pixel spatial Attention Module,PAM),来获得空间像素具有较强语义信息的特征图像.然后通过对PAM的输出进行建模提取,分别获得含有语义类别信息的边缘特征和小目标特征.最后,将特定的损失函数应用到语义分割训练中,并将多种特征进行融合,经过反复的监督学习和训练校正,可以在不影响其他类别性能的情况下提高边缘和小目标分割的性能.在Cityscapes,VOC2012,ADE20K和Camvid基线数据集上的实验表明,该算法与先进的图像分割算法相比,在小目标分割、边缘特征增强和内轮廓噪声减少等方面,其性能和效果都有明显提高,分割精度提高了2个百分点.

关键词: 场景分割, 小目标特征增强, 注意力模块, 建模

Abstract:

We have to face the challenge of missing small targets and severe edge noise in semantic segmentation. The existing semantic segmentation algorithms that enhance feature representation and optimize spatial details have difficulty to accurately segment the small targets and edges as the algorithms insufficiently gain detail information from tiny targets and semantic edges. This paper presents a tiny target feature enhancement algorithm for semantic segmentation. Specifically, a pixel spatial attention module(PAM) is designed to obtain strong semantic information from low-level pixel space. Semantic category information including edge features and tiny target features are obtained by modeling mask, respectively. A special loss function is designed for model training and the features gained by the model are fused with the features obtained from above way. Through edge feature enhancement, inner contour noise reduction, the segmentation performance of tiny target is improved while other segmentation categories are not degraded. Experimental results on Cityscapes, VOC2012, ADE20K and Camvid show that the proposed algorithm performance has been significantly improved by 2% in comparison with other state-of-the-art algorithms in the same scene.

Key words: scene segmentation, tiny target feature enhancement, attention module, modeling

中图分类号: