结合多种影像特征与CNN的城市建筑物提取

Urban Building Extraction by Combining Multiple Image Features and CNN

针对当前卷积神经网络对于城市建筑物纹理特征信息和多尺度信息利用的不足,提出了一种基于多种影像特征与卷积神经网络的城市建筑物提取方法,对结合尺度特征和纹理特征后的CNN模型的建筑物分类提取精度及其影响因素开展研究。方法采用局部二值模式来表达纹理特征,同时采用高斯金字塔提取多尺度特征,以此构建网络训练样本。基于此样本进行SegNet卷积网络训练,采用Softmax分类器完成建筑物粗提取,最后优化网络输出。研究表明,将纹理特征和尺度特征加入模型进行训练可以提高预测精度,其中精确率、召回率以及F1评分3个指标分别提升了8.01%、2.71%和4.98%。

Aiming at the current insufficient use of urban building texture feature information and multi-scale information by convolutional neural networks,a method for extracting urban buildings based on multiple image features and convolutional neural networks is proposed,which mainly combines scale features and texture features.The classification accuracy of building classification based on the CNN model and its influencing factors are studied.The method uses local binary model to express texture features,and uses Gaussian pyramid to extract multi-scale features to construct network training samples.Based on these samples,the SegNet convolutional network training is performed,the Softmax is constructed,rough extraction of buildings is completed,and finally the network output is optimized.Studies have shown that adding texture features and scales to the model for training can improve prediction accuracy.The accuracy,recall,and F1-Score are improved by 8.01%,2.71%,and 4.98%,respectively.