融合注意力机制和多尺度特征的图像语义分割  

Image Semantic Segmentation Based onFusing Attention Mechanism and Multi-scale Features

作　　者：姚庆安[1] 张鑫[1] 刘力鸣 冯云丛 金镇君 YAO Qing’an;ZHANG Xin;LIU Liming;FENG Yuncong;JIN Zhenjun(School of Computer Science and Engineering,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]长春工业大学计算机科学与工程学院,长春130012

出　　处：《吉林大学学报（理学版）》2022年第6期1383-1390,共8页Journal of Jilin University:Science Edition

基　　金：吉林省教育厅“十三五”科学技术研究规划项目(批准号:JJKH20200678KJ);吉林省科技发展规划重点研发项目(批准号:20200401076GX);符号计算与知识工程教育部重点实验室开放基金(批准号:93K172020K05).

摘　　要：针对图像语义分割中多尺度类别下目标分割率低、图像上下文特征信息关联性较差的问题,提出一种融合注意力机制和多尺度特征的图像语义分割模型.该模型用改进的带孔空间金字塔池化增加多尺度目标的分割,用注意力细化模块捕获上下文信息以启发特征学习,并加入基于注意力机制的特征融合有针对性地监督重要通道特征的学习,引导高、低阶特征融合,以提高模型的泛化能力.通过在数据集Cityscapes上的仿真实验结果表明,该模型的平均交并比相比DeepLab v3+提升了1.14%,证明了该模型具有较好的鲁棒性.Aiming at the problems of low target segmentation rate and feeble correlation of image context feature information under multi-scale categories in image semantic segmentation,we proposed an image semantic segmentation model that fused attention mechanism and multi-scale features.The model used the improved atrous spatial pyramid pooling to increase the segmentation of multi-scale targets,used the attention refinement module to capture context information to guide feature learning,and added feature fusion based on attention mechanism to supervise the learning of important channel features,guide the fusion of high-order and low-order features,so as to improve the generalization capability of the model.The simulation results on the Cityscapes dataset show that the mean intersection over union of the model is 1.14%higher than that of DeepLab v3+,which proves that the model has good robustness.

关 键 词：多尺度特征 特征融合 注意力机制 语义分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]