耦合Mask R-CNN和注意力机制的建筑物提取及后处理策略

Coupling Mask R-CNN and Attention Mechanism for Building Extraction and Post-Processing Strategy

建筑物是城市的重要组成部分,基于高分辨率遥感数据建筑物提取,在土地利用、城市规划和防灾减灾等方面有重要意义。针对建筑物提取存在的问题,提出一种改进的Mask R-CNN建筑物实例分割模型。基于残差神经网络融合卷积注意力模型,构建了残差卷积注意力网络,改善了特征提取不充分问题。通过添加Dice Loss的方法,对损失函数进行了优化,进而对特征学习过程进行了优化。并引入Douglas–Peucker algorithm、Fine polygon regularization algorithm相结合的后处理策略,使建筑物轮廓更规则。实验结果表明:改进模型相比原模型的检测mAP值在Iou 0.5时提高了7.74%、在Iou 0.75时提高了7.57%,后处理策略在选定合适阈值优化掩膜后较原始模型的F1-Score值提高了6.01%。耦合Mask R-CNN和注意力机制的实例分割模型改善了小型建筑物误检漏检问题、建筑物分割边界粘连问题,提高了建筑物的分割精度;优化了建筑物后处理策略,提高了建筑物规则化程度。

Buildings are integral components of urban areas.Extracting buildings from high-resolution remote sensing data holds significant academic importance in areas such as land use analysis,urban planning,and disas⁃ter risk reduction.For the problems of building extraction,an improved Mask R-CNN building instance seg⁃mentation model is proposed.Based on the residual neural network fusion convolutional attention model,a re⁃sidual convolutional attention network is constructed to improve the problem of inadequate feature extraction.The loss function is optimized by adding the Dice Loss method,and then the feature learning process is opti⁃mized.And a post-processing strategy combining Douglas-Peucker algorithm and Fine polygon regularization algorithm is introduced to make the building contours more regular and smooth.The experimental results show that the improved model improves the detection mAP value by 7.74%at Iou 0.5 and 7.57%at Iou 0.75 com⁃pared with the original model,and the post-processing strategy improves the F1-Score value by 6.01%com⁃pared with the original model after selecting the appropriate threshold to optimize the mask.The instance seg⁃mentation model coupled with Mask R-CNN and attention mechanism improves the small building misdetection and omission problem,building segmentation boundary adhesion problem,and building segmentation accuracy;building post-processing strategy,improves building regularization.