摘要
语义分割任务是对图像中的物体按照类别进行像素级别的预测,其难点在于在保留足够空间信息的同时获取足够的上下文信息。为解决这一问题,文中提出了全局双边网络语义分割算法。该算法将大尺度卷积核融入BiSeNet网络中,在BiSeNet网络原有的空间路径和上下文路径两条分支的基础上增加全局路径分支,使网络能够捕获更多的上下文信息,同时提出将BiSeNet网络中的注意力优化模块和特征融合模块中的全局池化模块替换为全局卷积模块,进一步提高了网络获取上下文信息的能力,从而使预测结果更加准确。实验结果表明,该算法在Cityscapes数据集上将交并比(MIoU)指标提高了0.84%,获得了优于BiSeNet网络的表现。
The task of semantic segmentation is to predict the obJects according to the category at the pixel level.The difficulty lies in retaining enough spatial information and obtaining enough context information.In order to solve this problem,this paper proposes a global bilateral network semantic segmentation algorithm.In this algorithm,the large-scale convolution kernel is integrated into the BiSeNet Network,and the global path branches are added to the original spatial path and context path of the BiSeNet Network,so that the network can capture more context information.At the same time,the global pooling module in the attention optimization module and feature fusion module is replaced by the global convolution module to further improve the network acquisition.The experimental results show that the algorithm improves the MIoU index by 0.84%on Cityscaps dataset,and achieves better performance than BiSeNet Network.
作者
任天赐
黄向生
丁伟利
安重阳
翟鹏博
REN Tian-ci;HUANG Xiang-sheng;DING Wei-li;AN Chong-yang;ZHAI Peng-bo(Institute of Electrical Engineering,Yanshan University,Qinhuangdao,Hebei 066000,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;Institute of Microelectronics,Chinese Academy of Sciences,Beijing 100029,China)
出处
《计算机科学》
CSCD
北大核心
2020年第S01期161-165,共5页
Computer Science
基金
国家重点研发计划(2018YFB1308302,2018YFB1308300)
复杂物体表面纹理获取和三维重建的关键技术研究项目(61573356)。
关键词
语义分割
双边分割网络
全局卷积网络
Semantic segmentation
Bilateral segmentation network
Global convolutional network
