利用深度残差网络的遥感影像建筑物提取

Building Extraction in Remote Sensing Imagery Based on Deep Residual Network

高分辨率遥感影像中建筑物的提取技术一直是遥感领域的研究热点。针对传统方法需要人工选取特征的缺点,提出一种结合深度残差网络结构和金字塔式层级连接的高分辨率遥感影像建筑物提取方法。首先对影像进行多尺度扩充,保证网络能够探测不同尺度建筑物特征;其次利用新提出的卷积神经网络训练模型,提取建筑物的像素级特征信息;然后对预测结果进行多模型集成计算,降低随机误差;最后对预测概率图选取合适的阈值,进行过滤去除椒盐噪声,利用形态学运算对结果后处理,保证建筑物完整,边界平滑。实验表明,相比于其他网络结构,所提网络结构的建筑物提取精度更高。

Building extraction from high resolution remote sensing imagery has always been a research hotspot.According to the shortcomings of traditional methods that rely on the handcrafted features,a method for building extraction from high resolution remote sensing imagery is proposed by combining deep residual network architecture and pyramid level connection.Firstly,the multi-scale expansion of images is carried out to ensure that the network can learn the building features of different scales.Secondly,the newly proposed convolutional neural network is used to train model to extract the pixel-level information of the buildings.Then,the multimodel integration calculation is performed on the prediction result to reduce the random error.Finally,the appropriate threshold is selected to filter the prediction probability map to remove salt-and-pepper noise,and the result is post-processed by morphological operation to make the building complete and the boundary smooth.Experimental results show that,compared with the other network architecture,the proposed network architecture has higher accuracy in the building extraction.