基于门控卷积残差网络的卫星图像道路提取

针对遥感影像中道路信息容易受到建筑物、植被等非道路信息干扰的问题,提出了一种基于门控卷积残差网络的遥感影像道路提取模型.首先,该网络使用ResNet101作为网络的编码器,在使得网络足够深的同时,也保证了梯度信息的有效传导;其次,在中心部分使用ASPP多尺度特征提取模块,进一步挖掘特征图中给予的信息;最后,使用门控卷积替换普通的卷积层,它可以根据特征图中参数的重要性,自适应分配权重,作为网络的解码器部分.该方法在CVPR DeepGlobe 2018道路提取挑战赛的数据集上进行了验证,平均交并比、Dice相似系数、召回率分别达到70.20％、82.06％、82.21％,均超过该赛事冠军DlinkNet34,提升了道路提取的效果.

Road extraction from satellite image based on gated convolutional residual network

This paper proposed a remote sensing image road extraction model based on gated convolution residual network, which aiming at the problem that road information in remote sensing images was interfered by non-road information. Firstly, the network used ResNet101 as the network encoder, which ensured the effective transmission of gradient information while making the network deep enough. Secondly, it used the ASPP multi-scale feature extraction module in the central part to further mine the information given in the feature map. Finally, it replaced the ordinary convolution layer by the gated convolution layer, which could be used as the decoder part of the network by adaptively assigning weights according to the importance of parameters in the feature graph. It verified the method on the dataset of CVPR DeepGlobe 2018 road extraction challenge. And the average crossover ratio, Dice similarity coefficient, and recall rate reach 70.20%, 82.06%, and 82.21%, respectively, which are all higher than DLinkNet34, the champion of the competition, and this method improves the effect of road extraction.