基于注意力机制与DUsamplingU-Net网络的建筑物提取

Building Extraction Based on Attention Mechanism and DUsamplingU-Net Network

针对常用语义分割模型因神经元接受域大小固定导致特征提取不充分、双线性上采样无法精确恢复像素级预测等问题,提出基于卷积核注意力机制和DUsampling(data-dependent upsampling)的卷积神经网络改进算法。算法引入卷积核注意力机制,利用选择性卷积核提取图像整体与局部细节特征,加强网络对建筑物特征的学习能力;在合并特征图之前将待融合的特征向下采样到特征图最低分辨率,通过DUpsampling分割标签空间的冗余准确地恢复像素级的预测;最后,结合卷积核注意力机制与DUpsampling构建SD-Unet模型,并在WHU数据集上进行验证。结果表明,该模型在交并比、总体精度、精确度、召回率以及F1分数上达到最优,分别为76.25%、98.86%、86.13%、87.07%和85.85%。SD-Unet模型加强网络对建筑物特征的学习能力及准确恢复特征图像素级的预测,比传统网络模型具有更好的分类性能和分类准确率。

In order to solve the problems of common semantic segmentation models,such as insufficient feature extraction due to the fixed size of neuron acceptance domain,and the inability to accurately recover pixel-to-level prediction by bilinear upsampling,an improved algorithm of convolutional neural network based on convolutional kernel attention mechanism and DUsampling(data-dependent upsampling)is proposed.Firstly,by introducing the convolution kernel attention mechanism,the algorithm utilizes the selective convolution kernel to extract the image feature in whole and local details to enhance the learning ability of the network to the building features.Secondly,before merging the feature map,features to be merged are subsampled to the minimum resolution of the feature map,and the pixel level prediction is accurately recovered through the redundancy of the label space by DUpsampling.Finally,the SD-Unet model is constructed by the convolution kernel attention mechanism and DUsampling,and is verified on the WHU data set.The results show that this model achieves the optimal results in terms of intersection over union,overall accuracy,recall rate and F1 score,which are 76.25%,98.86%,86.13%,87.07%and 85.85%respectively.The SD-UNET model enhances the learning ability of the network to the building features and accurately recovers the prediction of the pixel level of the feature map,which has better classification performance and classification accuracy than the traditional network model.