融合全卷积网络与条件随机场的高光谱语义分割

Hyperspectral Semantic Segmentation Fusing Fully Convolutional Networks and Conditional Random Field

针对高光谱遥感影像分割精度不足与单一网络信息流尺度存在局限性的问题,提出了基于双流框架的高光谱分割算法。算法融合深度学习模式中的全卷积神经网络(full convolutional networks,FCN)和高效判别式概率模型条件随机场(conditional random field,CRF),形成了高光谱影像语义分割算法FCN-CRF。在预处理阶段增加PCA(principal component analysis)降维,上采样阶段使用了混合上采样MUS(mix up-sampling)模块,形成了双流框架FCN-CRF分割算法。经过高光谱数据集Pavia University和Indian Pines测试,结果显示,相较于其他一些分割算法,FCN-CRF精度最高,总体精度分别达到了99.01%和98.60%,其参数量较少,运行效率较高。在不同地物类型中,该算法针对人工建筑物分割效果较植被好,边界保持较好。

Aiming at the problems of insufficient segmentation accuracy of hyperspectral remote sensing images and limitations of the scale of a single network information flow,this study proposes a hyperspectral segmentation algorithm based on a dual-stream framework.It combines the full convolutional neural networks(FCN)and conditional random field(CRF),forming a hyperspectral image semantic segmentation algorithm FCN-CRF.The algorithm adds PCA(principal component analysis)dimension reduction in the pre-processing stage,and uses the mix up-sampling(MUS)module in the upsampling stage.The dual-stream framework FCN-CRF segmentation algorithm has been tested in hyperspectral datasets such as Pavia University and Indian Pines.The results show that compared with other segmentation algorithms,FCN-CRF has the highest accuracy,with an overall accuracy of 99.01%and 98.60%,respectively.The number of parameters is less,and the operation efficiency is higher.Among different types of ground objects,its algorithm is better than vegetation in segmentation of artificial buildings,and the effect of boundary preservation is better.