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.
Nighttime Image Semantic Segmentation Based on Image Style Confrontation and Two-way Category Optimization
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.
万方数据
维普
