Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (22): 213-222.DOI: 10.3778/j.issn.1002-8331.2206-0255

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Real-Time Segmentation Network Based on Attention and Edge Extraction Tasks

WEN Kai, YANG Yipeng, XIONG Junchen, WEI Shengnan   

  1. 1.School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
    2.Research Center of New Telecommunication Technology Applications, Chongqing University of Posts and Telecommunications, Chongqing 400065, China
  • Online:2023-11-15 Published:2023-11-15



  1. 1.重庆邮电大学 通信与信息工程学院,重庆 400065
    2.重庆邮电大学 通信新技术应用研究中心,重庆 400065

Abstract: Aiming at the problems of low segmentation accuracy of small object and blurring segmentation boundary, BiSeNet V3 is improved, and a real-time segmentation network based on global attention and edge extraction tasks is proposed. Firstly, combined with asymmetric convolution, the short-term dense concatenate module is designed for lightweight, and the global correlation between deep-layer features is enhanced through global attention module. Secondly, edge branching can effectively filter out the boundary independent information in semantic features through edge feature fuse module, and recover the detailed information in the decoding stage. Finally, the improved loss function ensures that the network can update parameters toward the direction conducive to small object segmentation. Comparative experiments on datasets and real environment prediction results show that the proposed network improves the boundary definition and small object segmentation accuracy, and still has high robustness in real scenes.

Key words: small object, boundary blur, global attention, edge extraction task

摘要: 针对小目标分割精度不高且分割边界模糊等问题,对BiSeNet V3进行改进,提出了一种基于全局注意力及边缘提取任务的实时分割网络。结合非对称卷积,对短期密集拼接模块做了轻量化处理,并通过全局注意力增强了特征的全局相关性。边缘分支可通过边缘特征融合模块有效滤除语义特征中与边界无关的信息,并在解码阶段恢复损失的细节信息。改进的损失函数保证了网络能够向着利于小目标分割的方向更新参数。在数据集上的结果及真实环境预测表明,所提网络改善了边界清晰度及小目标分割精度,并在真实场景下仍具有较高鲁棒性。

关键词: 小目标, 边界模糊, 全局注意力, 边缘提取任务