数据安全视角下图像语义分割研究
Data security perspective on image semantic segmentation research
龚保全 1曾万康 1李绍毅 1彭熙1
作者信息
摘要
[目的/意义]在边缘场景下,实时图像的数据安全非常重要.同时,边缘计算设备为适应实时场景的需求,对图像语义分割提出新的要求,亟需提高模型运行速度与识别精度.[方法/过程]基于Jetson Nano分析了边缘设备数据安全的重要性和可行性,并改进PP-LiteSeg模型,引入高效多尺度注意力模块(EMA),采用融合上采样空洞空间卷积池化金字塔,提升了图像语义分割任务的性能.[结果/结论]在确保边缘设备数据安全的同时,改进模型较原有模型在Cityscapes数据集上平均交并比提升6.3%.
Abstract
[Purpose/Significance]Data security is very important for real-time images in edge scenarios,and edge computing devices pose new challenges for image semantic segmentation to meet the needs of real-time scenarios,which urgently requires improving the model speed and accuracy.[Method/Process]The research is based on Jetson Nano,and analyzes the importance and feasibility of data security for edge devices,and improves the PP-LiteSeg model,introduces an efficient multi-scale attention module(EMA),and adopts a fusion upsampling atrous spatial convolution pooling pyramid,which improves the performance of the image semantic segmentation task.[Results/Conclusion]While ensuring data security for edge devices,the improved model increases the mean IoU by 6.3%compared to the original model on the Cityscapes dataset.
关键词
数据安全/Jetson/Nano/高效多尺度注意力/空洞卷积/特征融合Key words
jetson nano/efficient multiscale attention/dilated convolution/feature fusion引用本文复制引用
出版年
2024
NETL
NSTL
万方数据