In this thesis a pipeline for the automated transfer of handwritten annotations from floor plan scans into the according BIM model is proposed. The main part of this thesis focuses on developing a suitable approach for realising the first step of translating pixel based, annotated floor plans into their valid digital representations. Related work proposes a deep learning based Generative Adversarial Network (GAN) for similar image to image translation tasks requiring a comparably small dataset for good prediction results. In combination with a suitable hyperparameter optimization based on educated guessing, this GAN architecture is shown to be applicable for the given task. For training and testing the prototype model a dataset comprising 350 unique real world floor plan samples is created by simulating handwritten annotations manually using a digital pen. Visual inspection of the results shows a proof of concept for the desired automated translation of manual changes within pixel-based floor plan images. An extensive and diverse dataset, adequate post processing and a suitable embedding of the presented approach into prevalent construction software applications are further steps for establishing a commercial, deep learning based routine for the automated integration of manual changes in printed floor plans into the corresponding digital BIM model.     «

In this thesis a pipeline for the automated transfer of handwritten annotations from floor plan scans into the according BIM model is proposed. The main part of this thesis focuses on developing a suitable approach for realising the first step of translating pixel based, annotated floor plans into their valid digital representations. Related work proposes a deep learning based Generative Adversarial Network (GAN) for similar image to image translation tasks requiring a comparably small dataset f...     »