改进SOLOv2的非结构化道路图像实例分割
Improved segmentation of unstructured road image instances in SOLOv2
宋亮 1谷玉海 1黄佳伟1
作者信息
- 1. 北京信息科技大学现代测控技术教育部重点实验室,北京 100192;北京信息科技大学机电工程学院,北京 100192
- 折叠
摘要
针对非结构化道路图像中多个目标重叠且间尺度差异较大等,容易漏检或错检,分割精度不佳等问题.提出一种改进的SOLOv2实例分割算法.首先在特征金字塔结构添加自底而上增强路径以减少特征传递过程的损失,其次使用双重注意力指导特征选择,自适应地选择重要特征,抑制冗余信息,提升细节特征的提取能力,增强类别分支和掩膜分支的特征表示,进而提高掩膜预测的准确率.此外,对非结构化道路图像数据集进行预处理操作,提高模型的泛化能力.实验结果表明,本方法对实例边界把控更为精准,对比SOLOv2和Mask-RCNN平均精度分别提高2.0％、2.2％,检测帧率提高到6.1帧/s,在不同环境下均具有良好的分割性能.
Abstract
In view of the overlapping multiple targets in unstructured road images and the large difference in scale,it is easy to miss or misdetect and poor segmentation accuracy.An improved SOLOv2 instance segmentation algorithm is proposed.Firstly,a bottom-up enhancement path is added to the feature pyramid structure to reduce the loss of fea-ture transfer process,and secondly,dual attention is used to guide feature selection,adaptive selection of important features,suppression of redundant information,improvement of the extraction ability of detailed features,and enhance-ment of feature representation of category branches and mask branches,so as to improve the accuracy of mask predic-tion.In addition,the unstructured road image dataset is preprocessed to improve the generalization ability of the mod-el.The experimental results show that the proposed method is more accurate in controlling the instance boundary,and the average accuracy of SOLOv2 and Mask-RCNN is increased by 2.0％and 2.2％,respectively,and the detection frame rate is increased to 6.1 frames/s,which has good segmentation performance in different environments.
关键词
实例分割/非结构化道路/注意力机制/深度学习Key words
instance segmentation/unstructured roads/attention mechanisms/deep learning引用本文复制引用
基金项目
北京市科技委促进高校内涵发展学科建设专项(5112011015)
出版年
2024
NETL
NSTL
万方数据