Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (18): 268-276.DOI: 10.3778/j.issn.1002-8331.2201-0412

• Graphics and Image Processing • Previous Articles     Next Articles

Multi-Feature Fusion for Camouflaged Object Segmentation

FU Bingyang, CAO Tieyong, ZHENG Yunfei, FANG Zheng, WANG Yang, WANG Yekui   

  1. 1.Institute of Command-and-Control Engineering, Army Engineering University of PLA, Nanjing 210007, China
    2.Firepower Department, The Army Artillery and Defense Academy of PLA, Nanjing 211100, China
    3.The Key Laboratory of Polarization Imaging Detection Technology of Anhui Province, Hefei 230031, China
  • Online:2022-09-15 Published:2022-09-15



  1. 1.陆军工程大学 指挥控制工程学院,南京 210007
    2.陆军炮兵防空兵学院南京校区 火力系,南京 211100
    3.安徽省偏振成像与探测重点实验室,合肥 230031

Abstract: In the field of camouflaged object segmentation, how to extract high-resolution semantic features from a depth model is the key to constructing a target segmentation model. In order to better solve this problem, a new camouflage target segmentation method based on multi-level feature fusion is proposed. A multi-stage gate control module is introduced to selectively fuse the multi-stage middle layer features of Res2Net-50, which can effectively filter the interference information of each level feature map during the feature encoding process. And in decoding, the self-interaction residual module has been used to drive the cross-fusion of encoding features of different scales, which guarantees the obtaining of more accurate target representation information. In addition, this paper combines cross entropy loss and Dice loss as a joint loss function to help the model segment the camouflaged target more accurately. Experimental results show that the proposed model performs better than the other eight typical models in the complex background camouflage data set and three common natural camouflage datasets.

Key words: deep learning, camouflaged object segmentation, feature fusion, gating mechanism, multiscale feature

摘要: 在伪装目标分割任务中,如何提取深度模型下高分辨率的目标语义特征是构建目标分割模型的关键。针对此问题,提出了一种基于多级特征融合的伪装目标分割方法。在特征编码过程中,引入多级门控模块对Res2Net-50的多级中间层特征进行选择性融合,有效过滤各级特征图的干扰信息;在解码过程中,通过自交互残差模块驱动不同尺度的编码特征实现交叉融合,获得更准确的目标表示信息。此外,在交叉熵损失的基础上加入Dice损失形成联合损失函数,帮助模型更精准地分割伪装目标。实验结果证明,在背景复杂的迷彩伪装数据集以及三个常用自然伪装数据集上,相比其他典型模型,该模型表现出更好的分割效果。

关键词: 深度学习, 伪装目标分割, 特征融合, 门控机制, 多尺度特征