基于余弦相似度的铸件浇冒口检测研究

Casting Riser Detection Based on Cosine Similarity

提出了自动化、智能化的铸件毛坯打磨切割工艺方案,对铸件浇冒口的分割检测方法进行了研究。基于RandLA-Net进行改进,提出了铸件点云分割网络FLA-Net,针对网络中仅仅考虑了距离信息而忽略了方向信息的不足,提出了一种结合余弦相似度的ACSE编码模块,提升了铸件浇冒口的分割精度。针对点云的无序性,参考PointCNN分割框架,使用X空间变换矩阵,使网络具有空间转换不变性,提升网络表达能力与语义分割精度。结果表明,对比经典点云语义分割框架,FLA-Net的效果更精确,具有更好的工程应用前景。

The automatic and intelligent cutting process scheme for casting blank grinding was provided,and the segmentation and detection method of casting riser was investigated.FLA-Net,a point cloud network,was put forward based on the improvement of RandLA-NET,and an ACSE coding module combining cosine similarity was proposed to improve the segmentation accuracy of casting riser aiming at the issues that only the distance information is considered and the direction information is ignored in the network.In view of the disorder of point cloud,X space transformation matrix is used referring to the PointCNN segmentation framework,so that the network has space transformation invariance,and the network expression ability and semantic segmentation accuracy are improved.Finally,the experimental results demonstrate that the FLA-Net is more accurate and has a better prospect of engineering application compared with the classical point cloud semantic segmentation framework.