A-LinkNet:注意力与空间信息融合的语义分割网络  

A-LinkNet:semantic segmentation network based on attention and spatial information fusion

作　　者：杜敏敏 司马海峰[1] DU Min-min;SIMA Hai-feng(School of Computer Science and Technology,Henan Polytechnic University,Jiaozuo 454000,China)

机构地区：[1]河南理工大学计算机学院,河南焦作454000

出　　处：《液晶与显示》2022年第9期1199-1208,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.61602157);河南省科技攻关项目(No.202102210167)。

摘　　要：针对道路图像语义分割中上下文信息不足以及空间细节信息易丢失等问题,本文提出一种基于LinkNet模型的实时分割方法。首先,在编码区域构建一种新的注意力机制,捕获道路图像的位置以及通道依赖,增加目标特征的提取能力。然后,在中心区域引入空洞空间金字塔池化模型,在不影响图像分辨率的情况下捕获更加丰富的多尺度特征。在通用数据库上的实验结果表明,所提方法在Cityscapes数据集上MIoU达到了64.78%,与LinkNet相比较,提高了5.01%,同时对于细小目标物体以及边界分割视觉效果有明显的改善,分割准确率获得了较大提升。To address insufficient contextual information and partial loss of spatial detail information in semantic segmentation of road images,a real-time segmentation method is proposed based on LinkNet.Firstly,a new attention mechanism is constructed in the encoding to capture the location and channel dependence of road images to increase the extraction capability of target features.Then,an atrous spatial pyramid pooling is introduced in the central region to capture richer multi-scale features without affecting image resolution.The experimental results on the general database show that the proposed method achieves 64.78%MIoU on the Cityscapes dataset,which is 5.01%higher in comparison with LinkNet.And it can significantly improve the visual effect of fine target objects and boundary segmentation,and the segmentation accuracy has been greatly improved.

关 键 词：语义分割 注意力机制 空洞空间金字塔池化 LinkNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]