改进SegNet网络对遥感图像的语义分割

Semantic Segmentation of Remote Sensing Images Based on Improved SegNet Network

为了更好地对包含丰富地表信息的高分辨率遥感图像进行语义分割,基于SegNet网络,结合金字塔池化(PPM),提出了P-SegNet网络,该网络能更好地对遥感图像的全局特征进行提取,分割效果优于改进前的SegNet网络.在训练集数量为120000张、迭代次数为20次的情况下,P-SegNet网络训练集Accuracy高于SegNet网络0.54%,达到96.36%,同时Loss相比Seg⁃Net网络减少了0.02,达到0.08.

In order to better semantically segment the high-resolution remote sensing image which contains rich surface informa⁃tion,P-SegNet network is proposed based on SegNet network and pyramid pooling(PPM).This network can better extract the global features of remote sensing image,and the segmentation effect is better than the unimproved SegNet network.When the number of training sets is 120000 and the number of iterations is 20,the accuracy of P-SegNet network training set is 0.54%higher than that of SegNet network,reaching 96.36%.Meanwhile,loss is 0.02 less than that of SegNet network,reaching 0.08.