预指导的多阶段特征融合的图像语义分割网络  

Image semantic segmentation network of pre-guidanced multi-stage feature fusion

作　　者：王燕[1] 范向辉 王丽康 Wang Yan;Fan Xianghui;Wang Likang(School of Computer&Communication,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学计算机通信学院,兰州730050

出　　处：《计算机应用研究》2024年第3期951-955,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61863025)。

摘　　要：针对目前语义分割对图像边缘和小物体不能进行精确识别,以及简单融合多阶段特征会造成信息冗余、混杂不清等问题,提出了一个预指导的多阶段特征融合的网络(pre-guidanced multi-stage feature fusion network,PGMFFNet),PGMFFNet采用编解码器结构,编码阶段利用预指导模块对各阶段信息进行指导,增强各阶段特征之间的联系,解决各阶段特征在后续融合过程中产生的语义混杂问题。在解码阶段,利用多路径金字塔上采样模块融合高级语义特征,然后使用改进的密集空洞空间金字塔池化模块对融合后的特征进一步扩大感受野,最后将高低层次的特征信息融合,使得对小物体的分割效果更优。PGMFFNet在CityScapes公开数据集上进行了验证,得到了78.38%的平均交并比(mean intersection over union,MIoU),分割效果较好。In view of the current semantic segmentation can not accurately identify image edges and small objects,and simple fusion of multi-stage features will cause information redundancy,confusion and other problems,this paper proposed a pre-guidanced multi-stage feature fusion network(PGMFFNet).PGMFFNet employed a encoder-decoder structure,at the encoder stage,which used a pre-guidance module to guide the information in each stage.Strengthened the relationship between the features of each stage,and solved the semantic confounding problems in the subsequent fusion process of the features of each stage.At the decoder stage,which used the multi-path up-pyramid sampling module to fuse high-level semantic features,and then used the improved dense void space pyramid pool module to further expand the sensory field of the fused features,and finally fused the feature information of high and low levels to make the segmentation effect of small objects better.This paper verified PGMFFNet on CityScapes open data set,and the mean intersection over union(MIoU)obtained to 78.38%,showing good segmentation effect.

关 键 词：语义分割 编解码器 预指导 金字塔 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]