Semantic Segmentation in Architectural Floor Plans for Detecting Walls and Doors

When architectural floor plans are rasterized, they lose valuable information, limiting human post-processing. To recover lost information, a process known as vectorization should be applied. Several studies have been concerned with vectorization. Even though previous methods do not perform well on images with various graphical styles and a large amount of noise, an approach based on deep learning has achieved exceptional results. The aim of this study is to segment walls and doors from architectural floor plan images using deep neural networks and construct a dataset with ground truth for wall segmentation.