《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (7): 2166-2172.DOI: 10.11772/j.issn.1001-9081.2022060933

• 人工智能 • 上一篇    下一篇

基于通道注意力和边缘融合的伪装目标分割方法

詹春兰1, 王安志1(), 王明辉2   

  1. 1.贵州师范大学 大数据与计算机科学学院,贵阳 550025
    2.四川大学 计算机学院,成都 610065
  • 收稿日期:2022-06-28 修回日期:2022-08-30 接受日期:2022-09-01 发布日期:2022-09-13 出版日期:2023-07-10
  • 通讯作者: 王安志
  • 作者简介:詹春兰(2000—),女,贵州毕节人,CCF会员,主要研究方向:伪装目标检测;
    王安志(1986—),男,贵州石阡人,副教授,博士,CCF会员,主要研究方向:深度学习、计算机视觉、数字图像处理;
    王明辉(1971—),男,四川成都人,教授,博士,主要研究方向:信息融合、医学影像处理、医疗大数据分析。
  • 基金资助:
    国家自然科学基金资助项目(62162013);国家级大学生创新创业训练计划项目(202210663045);贵州省大学生创新创业训练计划项目(S202110663028)

Camouflage object segmentation method based on channel attention and edge fusion

Chunlan ZHAN1, Anzhi WANG1(), Minghui WANG2   

  1. 1.School of Big Data and Computer Science,Guizhou Normal University,Guiyang Guizhou 550025,China
    2.College of Computer Science,Sichuan University,Chengdu Sichuan 610065,China
  • Received:2022-06-28 Revised:2022-08-30 Accepted:2022-09-01 Online:2022-09-13 Published:2023-07-10
  • Contact: Anzhi WANG
  • About author:ZHAN Chunlan, born in 2000. Her research interests include camouflage object detection.
    WANG Anzhi, born in 1986, Ph. D., associate professor. His research interests include deep learning, computer vision, digital image processing.
    WANG Minghui, born in 1971, Ph. D., professor. His research interests include information fusion, medical image processing, medical big data analysis.
  • Supported by:
    National Natural Science Foundation of China(62162013);National Innovation and Entrepreneurship Training Program for College Students(202210663045);Innovation and Entrepreneurship Training Program for College Students of Guizhou Province(S202110663028)

摘要:

伪装目标分割(COS)的目标是从背景中分离出隐藏的目标对象。近年来,基于卷积神经网络(CNN)的伪装目标检测(COD)发展迅速,然而仍存在无法从前/背景高度相似的场景中准确地检测出完整目标对象的问题。针对上述问题,提出一种基于通道注意力(CA)和边缘融合的COS方法CANet (Network based on Channel Attention and edge fusion),以得到伪装目标的边缘细节更清晰的完整分割结果。首先,引入压缩和激励(SE)注意力模块,以提取更丰富的高级语义特征;其次,提出一个边缘融合模块,抑制低级特征中的干扰,并充分利用图像的边缘细节信息;最后,设计了基于深度可分离卷积的通道注意力模块,以自上而下的方式逐步融合跨级的多尺度特征,进一步地提升检测精度和效率。在多个公开的COD数据集上的实验结果表明,相较于SINet (Search Identification Net)、TINet (Texture-aware Interactive guidance Network)和C2FNet (Context-aware Cross-level Fusion Network)等8种主流的方法,CANet表现更佳,且能够获取到丰富的伪装目标内部及边缘细节信息,而且在具有挑战性的COD10K数据集上结构度量指标相较于SINet提升了2.6个百分点。CANet性能优越,适用于医学上检测与人体组织相似的病灶区域、军事领域检测隐蔽目标等相关领域。

关键词: 伪装目标分割, 边缘融合, 压缩和激励注意力模块, 深度可分离卷积, 多尺度特征

Abstract:

The goal of Camouflage Object Segmentation (COS) is to detect hidden objects from the background. In recent years, Camouflage Object Detection (COD) based on Convolutional Neural Network (CNN) has developed rapidly, but there is still a problem that the complete object cannot be accurately detected in scenes with highly similar foreground/background. For the above problem, a COS method based on Channel Attention (CA) and edge fusion, called CANet (Network based on Channel Attention and edge fusion), was proposed to obtain a complete segmentation result with clearer edge details of camouflage objects. Firstly, the SE (Squeeze-and-Excitation) attention was introduced to extract richer high-level semantic features. Secondly, an edge fusion module was proposed to restrain interference in low-level features and make full use of edge details information of the image. Finally, a channel attention module based on depthwise separable convolution was designed to gradually integrate cross-level multi-scale features in a top-down manner, which further improved detection accuracy and efficiency. Experimental results on multiple public COD datasets show that compared to eight mainstream methods such as SINet (Search Identification Net), TINet (Texture-aware Interactive guidance Network) and C2FNet (Context-aware Cross-level Fusion Network), CANet performs better and can obtain rich camouflage objects’ internal and edge detail information. Among them, CANet improves the structure-measure index by 2.6 percentage points compared to SINet on the challenging COD10K dataset. CANet has superior performance and is suitable for medical detection of lesion areas similar to human tissue, military detection of hidden targets, and other related fields.

Key words: Camouflage Object Segmentation (COS), edge fusion, Squeeze-and-Excitation (SE) attention module, depthwise separable convolution, multi-scale feature

中图分类号: