融合注意力机制和多层次特征的语义分割方法  

Semantic Segmentation Method Integrating Attention Mechanism and Multi-level Features

作　　者：王威威 罗杨 李熹[1] 

机构地区：[1]广西民族大学人工智能学院,广西南宁530006

出　　处：《工业控制计算机》2025年第4期93-95,共3页Industrial Control Computer

摘　　要：当前存在的图像语义分割方法普遍存在无法充分利用全局信息和有效提取利用边缘信息的问题。针对以上问题,提出一种基于deeplabv3+的改进算法。该算法在编码端,加入一个non-local模块,以充分利用全局特征信息。在解码端加入CA(Coordinate Attention)模块提升目标边缘信息利用率。该方法在PASCAL VOC 2012数据集上进行实验,取得了77.4%的平均交并比结果。The existing image semantic segmentation methods generally have the problem that they cannot make full use of global information and effectively extract and utilize edge information.Aiming at the above problems,an improved algorithm based on deeplabv3+is proposed.In the encoder,a non-local module is added to make full use of the global feature information.The coordinate attention module is added to the decoder to improve the utilization rate of target edge information.The mIoU of this algorithm is 77.4%on PASCAL VOC 2012 dataset.

关 键 词：语义分割 多层次特征融合 全局注意力机制 NON-LOCAL CA 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]