This paper addresses the demand for automatic methods to manage and update public information stored in spatial databases, focusing on building footprint extraction and vectorization. Building footprints are vital for diverse applications like disaster management, urban monitoring, and cadaster updates. The paper presents an end-to-end workflow using deep learning, combining semantic segmentation with boundary regularization for building footprint extraction. Four convolutional neural network architectures are employed for binary semantic segmentation: U-Net, U-Net-Former, FT-UNet-Former, and DCSwin. The workflow begins with semantic segmentation using the trained models, followed by boundary regularization applied to the segmentation masks. The projectRegularization method combines semantic segmentation and boundary regularization through a generative adversarial network (GAN). The approach aims to create regularized building footprints with more consistent boundaries for cartographic and engineering applications. The workflow extends into developing an efficient vectorization methodology using open-source software solutions, aiming to make the results applicable in any GIS environment. The method is tested using the MapAI dataset and is intended to improve building footprint representations for practical use. The regularization and vectorization workflow is developed into a QGIS plugin, enhancing QGIS functionality. Overall, the paper seeks to advance convolutional neural network research for automatic building footprint extraction, contributing to open-source GIS software and improving the representation of building footprints.