改进SegNet+CRF高分辨率遥感影像建筑物提取方法

Building extraction from high remote sensing images based on Improved SegNet+CRF Method

摘要:

将传统的语义分割SegNet网络用于高分辨率遥感影像的建筑物提取时，分割的建筑物存在边界模糊、精度较低、错检漏检等问题。为了解决上述问题，提出一种改进SegNet网络+CRF语义分割方法。编码阶段的最低分辨率层引入空洞金字塔池化模型，通过并行的空洞卷积操作扩大特征提取的感受野；解码阶段构建特征金字塔实现特征多尺度融合，弥补上采样过程中丢失的特征信息；最后，预测图像送入全连接条件随机场模型进行后处理，优化提取的建筑物边缘。实验表明，相较于原SegNet网络，改进方法的建筑物提取像素精度、召回率、平均交并比分别提高了0.48%、1.29%、2.36%。

Abstract:

When using the traditional semantic segmentation SegNet network to extract buildings from high-resolution remote sensing images, there are some problems, such as fuzzy boundary, low accuracy, error detection and missing detection. In order to solve the above problems, an improved SegNet+CRF semantic segmentation method is proposed. The SegNet model is improved by adopting Atrous Spatial Pyramid Pooling(ASPP) in the encoding stage to extract the feature maps of different receptive fields of an image through multiple hole convolutions with designed expansion rates. In the decoding stage, construct Feature Pyramid Networks(FPN) to realize multi-scale feature fusion and reduce feature detail loss. Further, the prediction images are image-processed based on the fully connected conditional random field(CRF)model to optimize the building edges. Experimental results in test areas are evaluated quantitively and visionally, which show that the improved model has the higher accuracy than that of the original SegNet deep learning model, with the average pixel accuracy, recall and average cross-over ratio by 0.48% , 1.29% and 2.36% respectively. The improved method is able to acquire buildings with clear and accurate boundaries, and can be extended for recognition applications of in remote sending images for urban mapping, management and planning.