计算机科学
点云
人工智能
接头(建筑物)
嵌入
面子(社会学概念)
点(几何)
上下文图像分类
模式识别(心理学)
计算机视觉
图像(数学)
目标检测
计算
算法
数学
工程类
社会科学
几何学
社会学
建筑工程
作者
Zifan Shao,Kuangrong Hao,Bing Wei,Xue-song Tang
标识
DOI:10.1109/safeprocess52771.2021.9693672
摘要
In recent years, 3D classification has attracted a lot of attention, and many deep learning models have achieved good performance. In this paper, we propose a point cloud classification model for solder joint defect detection. Different from existing point clouds classification models such as MVCNN and PointNet, our model consumes six depth images. We present a standard CNN architecture learned to extract a global embedding of the 3D object from six depth images. We find that using depth images as input has great benefits. Because depth image is regular, we can simply use CNN to extract local features without any other measures. However, PointNet++ has to introduce KNN and farthest point sampling to abstract local features, which leads to expensive computation costs. So it will face a dilemma of either reducing the point cloud size in exchange for efficiency or increasing the number of the input points to achieve higher accuracy. Besides, depth images will not introduce geometry loss like rendered images used in MVCNN, and it is not necessary to deal with the uneven distribution problem of point clouds. Experiments show that our algorithm can achieve outstanding performance on fine-grained classification tasks such as solder joint defect detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI