传播深度和多尺度特征融合的遥感图像分割

Remote Sensing Image Segmentation Based on Propagation Depth and Multi-scale Feature Fusion

针对遥感图像因分辨率高、能见度低、物体种类多和光照不充足等因素导致遥感图像分割准确率低的问题,搭建一种传播深度和多尺度特征融合的端到端遥感图像分割网络。利用传播深度特征融合保留遥感图像的浅层特征信息,利用浅层特征指导深层特征的提取,在获得高级语义信息的同时没有过度损失浅层的位置特征信息情况下,通过改善并行空洞卷积模块的相关设置提取更多的多尺度信息,从而提高遥感图像分割的准确率。在Satellite datasetⅠ(global cities)数据集实验,相较于DeepLabv3网络所提出网络的准确率、精确率、召回率和F1值,分别提升了3.97%,12.95%,4.85%和13.23%,达到了88.79%,74.35%,85.62%和79.95%。实验结果表明,该网络能够有效地提高遥感图像分割的精确性。

In view of the low accuracy of remote sensing image segmentation due to high resolution,low visibility,many kinds of objects and insufficient lighting,an end-to-end remote sensing image segmentation network based on the fusion of propagation depth and multi-scale features is constructed.The shallow feature information of remote sensing image is preserved by using the propagation depth feature fusion,and the shallow feature is used to guide the extraction of deep feature,and in the case of obtaining advanced semantic information without excessive loss of shallow position feature information,more multi-scale information is extracted by improving the relevant settings of parallel hollow convolution module to improve the accuracy of remote sensing image segmentation.In the Satellite dataset I(global cities),compared with the accuracy,precision,recall rate and F1 values of the DeepLabv3 network,the accuracy,precision,recall rate and F1 values are increased by 3.97%,12.95%,4.85%and 13.23%respectively,to 88.79%,74.35%,85.62%and 79.95%.Experimental results show that the network can effectively improve the accuracy of remote sensing image segmentation.