文献类型：期刊文献

中文题名：基于改进BiSeNet的语义分割算法

英文题名：Semantic segmentation algorithm based on improved BiSeNet

作者：鹿鑫[1];杜煜[2];陈泽宇[1];徐世杰[1]

第一作者：鹿鑫

机构：[1]北京联合大学智慧城市学院,北京100101;[2]北京联合大学机器人学院,北京100101

第一机构：北京联合大学智慧城市学院

年份：2023

卷号：42

期号：7

起止页码：136-139

中文期刊名：传感器与微系统

外文期刊名：Transducer and Microsystem Technologies

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金青年科学基金资助项目(61803034);北京联合大学研究生科研创新资助项目(YZ2020K001)。

语种：中文

中文关键词：双路径;特征融合;梯度信息;语义分割;无人驾驶

外文关键词：dual path;feature fusion;gradient information;semantic segmentation;unmanned driving

摘要：语义分割在无人驾驶环境感知方面有重要的研究意义,需要同时兼顾语义和空间信息。针对提取语义信息时空间信息丢失不可恢复的问题,提出一种基于改进BiSeNet的语义分割算法。首先,针对空间路径信息冗余的问题,将输入图像转换为灰度图,并通过Sobel算子提取梯度信息后,采用卷积结构提取空间信息;然后,对两分支不同分辨率特征图进行逐步融合上采样。在Cityscapes数据集上对算法进行验证,分割精度达到65.79%,相比BiSeNet提高了3.11%。
Semantic segmentation has important research significance in perception of unmanned driving environment,and it needs to take into account both semantic and spatial information.Therefore,aiming at the problem of unrecoverable loss of spatial information when extracting semantic information,a semantic segmentation algorithm based on improved BiSeNet is proposed.Firstly,aiming at the problem of spatial path information redundancy,the input image is converted into grayscale image,and the gradient information is extracted through the Sobel operator,and then the convolution structure is used to extract the spatial information.Then,the feature maps of the two branches with different resolutions are gradually fused and upsampled.The algorithm is verified on the Cityscapes dataset,and the segmentation precision reaches 65.79%,which is 3.11%higher than BiSeNet.