基于门控注意力融合的电子元器件点云联合分割

Joint segmentation of point clouds of electronic componentsbased on gated attention fusion

为解决现有联合分割算法点云分割过程中共享编码模块与多支线融合模块存在潜在特征冲突的问题,提出了基于门控注意力融合的电子元器件点云联合分割方法,在共享编码模块浅层特征层引入门控传播模块(GCM),通过权重自学习的方式分别强化各支线特征,抑制无关信息流动;在多支线融合模块提出联合注意模块(CAM),通过捕获空间及通道特征权重,进而加权完成多支线特征软融合.在S3DIS数据集上,所提方法的语义分割平均交并比(mIoU)达56.2%、平均准确率(mAcc)达64.4%;实例分割平均实例覆盖率(mCov)达51.4%、平均实例精度(mPrec)达56.5%.在电子元器件数据集上,上述指标分别达84.5%、91.6%、85.9%及86.5%.实验表明,所提方法有效解决了特征冲突问题,分割精度高于主流联合分割算法.

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.