Joint segmentation of point clouds of electronic components based on gated attention fusion
In order to solve the problem of potential feature conflicts between the shared coding module and the multi-branch fusion module in the point cloud segmentation process of the existing joint segmentation algorithm,a joint point cloud segmentation method for electronic components based on gated attention fusion is proposed.The layer feature layer introduces the gated propagation module(GCM),which strengthens the characteristics of each branch through weight self-learning,and suppresses the flow of irrelevant information;in the multi-branch fusion module,a joint attention module(CAM)is proposed,which captures space and channel feature weights,then weighted to complete the soft fusion of multi-branch features.On the S3DIS dataset,the average intersection-over-union ratio(mIoU)of the proposed method reaches 56.2％,and the average accuracy rate(mAcc)reaches 64.4％;the average instance coverage(mCov)of instance segmentation reaches 51.4％,and the average instance accuracy(mPrec)reached 56.5％.On the electronic component data set,the above indicators reach 84.5％,91.6％,85.9％and 86.5％,respectively.Experiments show that the proposed method effectively solves the feature conflict problem,and the segmentation accuracy is higher than the mainstream joint segmentation algorithm.
point cloud joint segmentationattention mechanismelectronic component
NETL
NSTL
万方数据
维普
