系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (9): 2707-2715.doi: 10.12305/j.issn.1001-506X.2022.09.02

• 电子技术 • 上一篇    下一篇

一种基于定位和非对称补偿的伪装目标分割方法

徐义飞1,2, 李晓冬2, 李新德1,3,*   

  1. 1. 东南大学自动化学院, 江苏 南京 210096
    2. 信息系统工程重点实验室, 江苏 南京 210000
    3. 南京数学应用中心, 江苏 南京 211135
  • 收稿日期:2021-11-19 出版日期:2022-09-01 发布日期:2022-09-09
  • 通讯作者: 李新德
  • 作者简介:徐义飞(1996—), 男, 硕士研究生, 主要研究方向为目标检测、语义分割|李晓冬(1986—), 男, 高级工程师, 硕士, 主要研究方向为智能信息处理|李新德(1975—), 男, 教授, 博士, 主要研究方向为人工智能、智能机器人、机器视觉感知
  • 基金资助:
    信息系统工程重点实验室开放基金(05202003)

A method of camouflaged object segmentation with locating and asymmetric compensation

Yifei XU1,2, Xiaodong LI2, Xinde LI1,3,*   

  1. 1. School of Automation, Southeast University, Nanjing 210096, China
    2. Science and Technology on Information System Engineering Laboratory, Nanjing 210000, China
    3. Nanjing Center for Applied Mathematics, Nanjing 211135, China
  • Received:2021-11-19 Online:2022-09-01 Published:2022-09-09
  • Contact: Xinde LI

摘要:

伪装是欺骗观察者感知系统的一种手段, 善于伪装的个体在纹理特征上与背景具有高度的相似性。为解决前景与背景因相似而导致的像素归属歧义, 提出一种基于定位和补偿网络(locating and compensation network, LCNet)的伪装目标分割网络。该方法效仿了捕食者从搜索→确立→聚焦的寻猎过程, 涵盖双主干网的强感知提取、定位模块的双注意力以及级联的非对称补偿模块的细化像素模糊。实验表明, 在4种评价指标下, LCNet在3个具有挑战的伪装数据集上都显著优于现有的6种最新模型, 具有较高分割性能。

关键词: 纹理伪装目标, 非对称注意力补偿, 双注意力定位, 双主干网

Abstract:

Camouflage is the instinct to deceive the perception of the observer, which presents the high similarity of the textural characteristics with the surroundings. In order to address the ambiguous regions generated by the similarity between background and foreground, this paper proposes a camouflaged object segmentation network based on locating and compensation network (LCNet). Inspired by predator process: Search→Establish→Focusing, the paradigm of the proposed method is achieved via dual-backbone with the strong sense of knowledge extraction, locating module with double attention, and cascading asymmetric compensation module with pixel refining. Experimental results have shown that the performances of LCNet are superior than six state-of-the-art models at the three major challenging camouflaged datasets in terms of four metrics, and the effectiveness of LCNet is demonstrated.

Key words: textural camouflaged object, asymmetric attention compensation, double attention locating, dual-backbone

中图分类号: