Improvement of DGCNN based on point attention module and generation of hot-rolled surface inspection dataset

The use of point clouds in computer graphics is increasing, providing a versatile geometric representation for various applications and serving as the main output of many 3D data acquisition tools. Despite the traditional use of hand-designed attributes on point clouds in graphics and vision fields, recent success of Convolutional Neural Networks (CNNs) in image analysis suggests that using CNN understandings to process point clouds holds significant promise. To address the lack of inherent topological information in point clouds, we enhanced the DGCNN network by introducing a lightweight attention module, named DGCNN-PA (Dynamic Graph CNN based on the Point Attention module). Additionally, creating extensive datasets is crucial to meet industry demand and support unsupervised macro-model training. We propose a methodology for generating three-dimensional point cloud using a two-dimensional segmentation model and relative depth estimation. Our model achieves an mIoU of 86.2 on the ShapeNetPart dataset and 64.2 on the hot-rolled steel Strip Surface Inspection dataset.