摘要
针对传统语义分割模型对监控场景下具有不同明亮程度图像分割效果不佳的问题,提出了一种融合多曝光图像的监控场景分割模型HDR-PSPNet.此模型通过融合多张不同曝光时间的图像实现数据增强,使用空洞卷积替代金字塔池化模块来保障特征图具有相同的分辨率,同时使用编解码结构增强图像底层特征的提取能力.通过基于自建新疆某水库监控图像数据集的实验结果可知,HDR-PSPNet相较于FCN、PSPNet和DeepLabv3,y MPA指标分别增加了5.5%,1.6%和1.0%,x MIoU指标分别增加了6.4%,2.9%和2.1%,表明其在多曝光时间图像监控场景的分割上HDR-PSPNet效果优于FCN、PSPNet和DeepLabv3网络.
In order to tackle the problem of poor segmentation effect on images with different brightness levels in surveillance scenes by traditional semantic segmentation models,a surveillance scene segmentation model based on multi-exposure image fusion,named HDR-PSP,is proposed.In this model,data enhancement is achieved by fusing multiple images with different exposure durations,dilated convolution is used instead of pyramid pooling to ensure the same resolution of feature maps,and the encoding-decoding structure is used to enhance the low-level feature extraction capability.The experiments are completed using self-labeled surveillance scene images of a reservoir in Xinjiang as the dataset.Compared with FCN,PSPNet and DeepLabv3 models,the y MPA indicator of HDR-PSPNet has increased by 5.5%,1.6%and 1.0%respectively,and the x MIoU indicator has increased by 6.4%,2.9%and 2.1%respectively.It shows that HDR-PSPNet can do better than FCN,PSPNet and DeepLabv3 on surveillance scene segmentation.
作者
乔金明
朱耀琴
QIAO Jin-ming;ZHU Yao-qin(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《东北师大学报（自然科学版）》
CAS
北大核心
2023年第3期75-84,共10页
Journal of Northeast Normal University(Natural Science Edition)
基金
国家自然科学基金资助项目(62172221).
作者简介
乔金明(1996-),男,硕士;通信作者:朱耀琴(1977-),女,博士,讲师,主要从事系统仿真与仿真决策研究.
