Deep Learning based Point Classification for Ship Hull Plate Model Reconstruction

Abstract

In the shipbuilding industry, shipyards generally estimate the production cost of a ship before the fabrication, so they can avoid unnecessary tasks to reduce production time and cost. The ship fabrication is composed of several steps, and recently robots and CNC machines usually aid most of them, which saves both time and cost. However, the forming process of curved hull plates still requires manual work, so it is time-consuming. Furthermore, only skillful workers participate in the formation as they determine the method of the forming process. In addition, forming time can vary for the same worker because the working environment or the worker’s condition can affect the forming process. Thus, various forming time or cost estimation research in the past first assume that the forming time depends on the target shape of curved hull plates, which means the forming time or cost increases as the shape complexity of the target curved hull plate increases. All methods may differ, but they all first extract geometric information from the shape of the target hull plates to determine forming time/cost. The problem is that all these studies assume that accurate 3D models of curved hull plates are already given, which most shipyards do not have. Most shipyards generally possess 2D plans, or even if they possess 3D models, most 3D hull models are not accurate for geometric information extraction. A framework of 3D model reconstruction method for curved hull plates is proposed to overcome this problem.

In this dissertation, the deep learning based Point Classification for Curved Ship hull plate model reconstruction is proposed. When a 3D point set from the target hull plate is given as input, the proposed method first removes duplicated points, and estimated index value for each point. Subsequently, the proposed method determined four corner points of rectangular shape based on the index value. After corner points are determined, the proposed method classified points into boundary or interior points based on the index value. To classify the points, the proposed method built the 4D point set by adding the index value, and used the existing deep learning network, PointNet. To train PointNet, the proposed method created synthetic hull plates with various shapes and grid sizes. Additionally, real hull plates are added to the training dataset for efficient training. After PointNet classified points into boundary or interior points, the proposed method moves to the final process. The proposed method reconstructed the four boundary curves based on classified points. Finally, a 3D model is obtained from the reconstructed boundary curves. The experiment results demonstrate that the proposed method can accurately reconstruct the 3D model from curved hull plates using only a sparse pointset since addition, the proposed method can accurately classify points into boundary or interior points for real hull plates compared to existing methods.