基于图像风格对抗和二重类别优化的夜间图像语义分割  

Nighttime Image Semantic Segmentation Based on Image Style Confrontation and Two-way Category Optimization

作　　者：周华平 李美光 ZHOU Huaping;LI Meiguang(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232000,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232000

出　　处：《湖北民族大学学报（自然科学版）》2024年第1期69-74,共6页Journal of Hubei Minzu University:Natural Science Edition

基　　金：安徽省重点研发计划国际科技合作专项基金资助项目(202004b11020029)。

摘　　要：为解决目前夜间图像语义分割中存在的语义信息传递丢失和不重视小频率类别问题,提出了基于图像风格对抗和二重类别优化网络架构(image style confrontation and two-category optimization network architecture,ITA)模型。首先,采用了对抗学习的架构令图像共享信息得到高效利用,使得语义信息的传输更加完整。然后,采用了二重类别指导策略(two-way category guidance,TCG)重新分配类别权重,引导模型更加关注小频率类别。最后,在黑暗苏黎世数据集上进行实验,平均交并比(mean intersection over union,MIoU)提高到了60.1%。另外,通过消融实验证明了每个模块的有效性。ITA模型能够较为准确地分割夜间道路图像,可供夜间自动驾驶任务借鉴。In order to solve the problem of semantic information transmission loss and lack of attention to small frequency categories in the semantic segmentation of nighttime images,an algorithm model based on image style confrontation and two-category optimization network architecture(ITA)was proposed.Firstly,the architecture of adversarial learning was adopted,so that the image sharing information could be used efficiently,and the transmission of semantic information was more complete.Then,the framework also adopted the two-way category guidance(TCG)strategy to reallocate the class weights,and the guidance model paid more attention to the small-frequency classes.Finally,the mean intersection over union(MIoU)in the dark Zurich dataset increased to 60.1%.Meanwhile,the effectiveness of each module was also demonstrated by ablation experiments.The ITA model framework could accurately segment the nighttime road image,which could provide reference for the night autonomous driving task.

关 键 词：夜间图像 语义分割 对抗网络 域自适应 图像对齐 数据平衡 风格转换 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]