Masterarbeit (laufend)

Lisa von Rössing

Symbol Detection in Technical Drawings: Independent of Drawing Style through Machine Learning

A so-called digital twin of a structure is used to optimize its construction, operation,
maintenance and energy usage under economic, environmental and safety aspects, among others.
But many structures exist just as physical assets, with analogue documentation such as construction drawings available only on paper. Therefore, construction drawings and other documents need to be digitized to make their information available for further processing.
For this, computer vision is utilized, which has advanced immensely with the usage of deep neural networks for tasks such as object detection. But neural networks consist of millions of parameters, meaning that effectively tuning them on the given task necessitates massive amounts of training data. To teach a new class to an already pre-trained model, thousands of high-quality labelled images are required.
Gathering and annotating them is time-consuming and laborious. Additionally, in specialized areas such as civil engineering, this annotation requires experts. Furthermore, gathering initial images for training is made difficult by e.g. security and privacy considerations, and finally, drawing styles and symbols are not standardized.
Few-shot detection (FSD) models detect new classes based on a limited number of examples, commonly zero to thirty instances. These models require neither a large amount of training data nor a significant amount of training time to detect novel classes, and therefore present a promising solution to data scarcity.
Because FSD models have not been tested on construction drawings, the thesis investigates the feasibility of minimizing training costs of symbol detection in architectural drawings by utilizing available FSD models when considering the examples of the CubiCasa5k image set and bridge construction plans.