摘要
随着城市道路变得越来越复杂,实现道路场景图像在像素级别上的分类成为自动驾驶研究的基础。道路场景中对于外界因素如光照、遮挡、尺寸会使得目标物体与图片上差异较大,对语义分割有着更高要求。选用基于金字塔池化模块(Pyramid Scene Parseing Network,PSPNet)语义分割算法进行道路场景分割,在进行特征提取时选择使用深度可分离卷积、残差边模块轻量型的MobileNetV2网络,得到特征图经过PSP模块时平均池化分成4个区域,区域大小为1×1、2×2、3×3以及6×6,将4个区域结果统一尺寸,并输入PSP模块特征图做堆叠,最终分割出每个像素代表类别。实验结果表明,使用PSPNet模型能够有效对道路场景进行语义分割。
As urban roads become more complex, achieving classification of road scene images at the pixel level becomes the basis for autonomous driving research. Road scenes have higher requirements for semantic segmentation for external factors such as illumination, occlusion, and size that can make the target object differ from the picture. The Pyramid Scene Parseing Network(PSPNet)semantic segmentation algorithm is chosen to segment the road scenes, and the MobileNetV2 network with separable convolutional depth and lightweight residual edge module is selected for feature extraction. The region sizes are 1’1, 2’2, 3’3, and 6’6. The results of the four regions are uniformly sized and input into the PSP module feature map to do stacking, and finally segment each pixel to represent a category. The experimental results show that the use of PSPNet model can effectively perform semantic segmentation of road scenes.
作者
张绪德
唐厚炳
ZHANG Xude;TANG Houbing(Kaili University,Kaili Guizhou 556011,China)
出处
《信息与电脑》
2022年第19期70-73,共4页
Information & Computer
基金
贵州省教育厅青年科技人才成长项目(项目编号:黔教合KY字〔2022〕079号)。
关键词
语义分割
平均池化
深度可分离卷积
semantic segmentation
average pooling
depth separable convolution
