在高分辨率遥感影像地类识别上,语义分割网络Deeplabv3+表现优异,但是所需参数非常多,训练时间久.遥感影像中的地类与普通RGB图片中的对象相比颗粒度非常大,其具有更显著的特征以及更少的类别,并不需要过深过大的网络.因此,文中提出了一种基于轻量级网络的语义分割模型(Thin-Deeplabv3+),对 Deeplabv3+的编码器进行了改进,利用轻量级膨胀网络(Light and Dilated Network,LDNet)提取输入遥感影像的特征,然后利用膨胀系数分别为 2、12、24 和 36 的空洞空间金字塔池化(Atrous Spatial Pyramid Pooling,ASPP)模块加强特征提取.在高分遥感影像数据集(Gaofen Image Dataset,GID)以及DeepGlobe土地覆盖分类挑战数据集(DeepGlobe Land Cover Classification Challenge,DLCCC)上的实验结果表明,Thin-Deeplabv3+的识别精度高于Deeplabv3+,并且所需参数仅约为Deeplabv3+的1/10.
Improved Deeplabv3+for land class identification in high resolution remote sensing images
The semantic segmentation network Deeplabv3+is adept at land class recognition for high-resolution remote sensing images,but it requires considerable parameters and costs extra time for training.Compared with the objects in ordinary RGB images,the earth class in remote sensing images is very grainy,with more significant features and fewer categories.It does not need a network too deep or too large.Therefore,this paper proposes a semantic segmentation model based on lightweight network(Thin-Deeplabv3+),which improves the Deeplabv3+encoder,and uses the light and dilated network(LDNet)to extract features from input remote sensing images.Then,the atrous spatial pyramid pooling(ASPP)module with expansion coefficients of(2,12,24,36)is used to enhance feature extraction.Finally,experiments are conducted on the Gaofen Image Dataset(GID)and the DeepGlobe Land Cover Classification Challenge(DLCCC),and the results show that the recognition accuracy of Thin-Deeplabv3+is higher than that of Deeplabv3+,and the required parameter number is only about one tenth of that of Deeplabv3+.
light and dilated network(LDNet)land class identificationremote sensing imageThin-Deeplabv3+
万方数据
维普
