Exploring the Capability of Deep Neural Network on Ship Sub-Assembly Weld Seam Recognition

To address the problem of robot weld process planning for complex ship sub-assembly work-pieces, this paper explores the capability of the deep neural network models on seam cloud point segmentation tasks. By studying and leveraging the performance of three popular point cloud segmentation models, we choose PointNet++ as the backbone and introduce three improvements to enhance its performance on the weld seam segmentation problem. First, a multi-perspective edge detection method is proposed to improve segmentation accuracy. Second, neighborhood sampling and fast farthest point sampling are introduced to improve the sampling efficiency. Finally, we use fewer convolution layers to reduce the time and space complexity and resolve the overfitting problem. A dataset of sub-assembly models containing 4 plate type classes and 5 seam type classes was established. The training and recognition performance of the advanced PointNet++ model was evaluated and compared with the traditional geometrical weld seam recognition method. The experimental results prove that the proposed method can adapt to 5 typical weld seam classes at a mIoU 83.38%, and it is robust to arbitrary work-piece layout positions, meanwhile, it reaches a speed at 15fps. All the above results demonstrate the great potential of deep neural networks on the robot weld applications for ship sub-assembly work-pieces.