Skip to Main content Skip to Navigation

Unsupervised and hybrid vectorization techniques for 3D reconstruction of engineering drawings

Salwan Alwan 1, 2 
1 Lab-STICC_M3 - Equipe Marine Mapping & Metrology
Lab-STICC - Laboratoire des sciences et techniques de l'information, de la communication et de la connaissance : UMR6285
Abstract : Computer-aided design technologies are highly improved during the last decade. This revolution created a gap between old technical drawings (handmade saved as raster images) and the new technical drawings (saved as vector data), which can be modified and saved easily. Moreover, this gap grows up with the rising of virtual and augmented reality applications. Different algorithms can reconstruct 3D models from vector technical drawings. Therefore, a computer-aided design system for converting raster images into vector data is needed more than ever. In this thesis, a preprocessing framework is proposed to prepare data for the vectorization process. The framework extracts graphical information from the engineering drawing template, separates different views using a clustering approach, and denoises each view separately by adopting a deep learning network. Moreover, the framework generates the skeleton of the image, whic his used in the vectorization process. Finally, the framework detects arrow heads where arrowheads can be lately a pattern to detect dimension sets. The genetic algorithm approach is adapted to vectorize technical drawings. This method tune geometric parameters based on the estimated width, which decreases the possibility of fragmenting primitives. The evaluation shows the robustness and effects of hyperparameters on the proposed unsupervised vectorization approach. A hybrid method is proposed to reduce the complexity of the vectorization problem. The supervised stage uses deep learning networks to segment the input image into different layers where each layer contains only one type of primitives (such as straight-line layer and circle layer). The unsupervised stage detects the primitive in each layer separately. The hybrid vectorization approach converts the problem from curve segmentation (in unsupervised vectorization approach) in to primitive detection. Moreover, the hybrid vectorization approach decreases the time complexity due to simultaneously detecting different types of primitives.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Monday, February 7, 2022 - 3:51:20 PM
Last modification on : Monday, April 4, 2022 - 9:28:32 AM
Long-term archiving on: : Sunday, May 8, 2022 - 6:52:21 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03560506, version 1


Salwan Alwan. Unsupervised and hybrid vectorization techniques for 3D reconstruction of engineering drawings. Computer Vision and Pattern Recognition [cs.CV]. Ecole nationale supérieure Mines-Télécom Atlantique, 2021. English. ⟨NNT : 2021IMTA0254⟩. ⟨tel-03560506⟩



Record views


Files downloads