PU learning-based recognition of structural elements in architectural floor plans

This work introduces a computational method for the recognition of structural elements in architectural floor plans. The proposed method requires minimal user interaction and is capable of effectively analysing floor plans in order to identify different types of structural elements in various notation styles. It employs feature extraction based on Haar kernels and PU learning, in order to retrieve image regions, which are similar to a user-defined query. Most importantly, apart from this user-defined query, the proposed method is not dependent on learning from labelled samples. Therefore, there is no need for laborious annotations to form large datasets in various notation styles. The experimental evaluation has been performed on a publicly available and diverse dataset of floor plans. The results show that the proposed method outperforms a state-of-the-art method, with respect to retrieval accuracy. Further experiments on additional floor plans of various notation styles, demonstrate its general applicability.