融合注意力和多尺度表达的机载激光点云分类

Attention mechanism and multi-scale feature fusion network for ALS point cloud classification

针对机载激光点云数据中存在空间分布不均匀和地物尺度不一的问题,提出了一种融合注意力机制和多尺度特征的机载激光点云分类深度学习方法(AMMSF-Net)。该方法建立了局部空间位置注意力层学习局部邻域上下文特征,增加注意力跳连机制将解码器和编码器中的特征进行动态融合并有效保留细节信息;解码器中的多尺度特征融合通过将不同尺度的特征进行级联输入到多层感知机和条件马尔可夫层得到最后的语义概率图,实现了不同尺度与不同层级特征图之间的相关,增强不同尺度目标的表达能力。在Vaihingen数据集中AMMSF-Net取得83.8%的总体精度和70.4%平均F1分数,在DFC3D数据集取得了95.4%总体精度和88.5%平均F1分数,对比其他模型该方法在两个数据集都取得了更好的精度,这表明AMMSF-Net能有效提高点云地物类别区分的能力。

Airborne Laser Scanning(ALS)point cloud classification is essential to extract geoinformation,but the uneven spatial distribution and scale variations between different categories bring challenges to the fine classification of point cloud data.In this paper,an attention mechanism and multi-scale feature fusion deep learing network(AMMSF-Net)for ALS point cloud classification was proposed.In the network a local spatial position attention layer was used to learn local contextual features;and an attention skip connection scheme was added to dynamic fusion the corresponding features among the encoder and decoder,which can retain detail features and contextual information.The multi-scale feature in the decoder fusion module obtained the final semantic probability map by concatenating the features at different scales into MLP(Multilayer Perceptron)and CML(Conditional Markov Layer),which can achieve the correlation of the feature maps between different scales and different levels,and can enhance the expression of targets at different scales.In two datasets experiment,AMMSF-Net achieveed 83.8%overall accuracy and 70.4%average F1 score in the Vaihingen dataset,and 95.4%overall accuracy and 88.5%average F1 score in the DFC3D dataset.Compared with other popular method,AMMSF-Net got higher classification accuracy in both datasets,which shows that AMMSF-Net can distinguish ground objects in point cloud effectively.