Machine Learning-based Segment and Object Detection for Engineering Applications

The digitization of legacy engineering drawings, created and stored on paper, is necessary to keep projects up to date and enable future alterations. Machine learning approaches emerge as tools to automate the digitization process by interpreting, classifying, and converting paper-based images into digital formats. This thesis investigates the application of various machine learning techniques to extract geometric and semantic information from railway engineering schematics and proposes a complete solution for their digitizing. The proposed method automates the detection and classification of individual diagram components by applying deep learning methods, segment detection, and clustering techniques. These components are then connected into graph structure using domain-specific ontology. The results demonstrate the method's ability to reconstruct engineering schematics automatically from paper to a vectorized drawing interface. Considering that rare symbols are present in these projects and are difficult to identify since there are not enough samples to train the model properly, generative models are used to augment the training dataset, improving its performance by using a larger number of analyzed samples. In future works, the graph structure created here can be used to validate the digitization of engineering drawings. The methods for digitizing presented here can be applied to other drawings and are a general solution that can be used for a variety of applications.