In this work we investigate the use of deep neural networks for object detection in floor plan images. Object detection is important for understanding floor plans and is a preliminary step for their conversion into other representations. In particular, we evaluate the use of object detection architectures, originally designed and trained to recognize objects in images, for recognizing furniture objects as well as doors and windows in floor plans. Even if the problem is somehow easier than the original one in the case of this research the datasets available are extremely small and therefore the training of deep architectures can be problematic. In addition to the use of object detection architectures for floor plan images, another contribution of this paper is the creation of two datasets that have been used for performing the experiments covering different types of floor plans with different peculiarities.

No abstract

In this paper we present a method for text recognition in floor plan images. In particular, we are concerned about locating, reading, and categorizing text inside floor plan images to obtain information about the building. Furthermore, the aim of this paper is to compare traditional text detection methods, based on image processing techniques, with recent approaches relying on convolutional neural networks. To improve results we combined several methods outperforming the original ones. Text regions are also classified in four semantic classes according to their purpose. Two datasets with different features, including quality and size, were considered in the experiments performed.

Digital reproductions of historical documents from Late Antiquity to early medieval Europe contain annotations in handwritten graphic symbols or signs. The study of such symbols may potentially reveal essential insights into the social and historical context. However, finding such symbols in handwritten documents is not an easy task, requiring the knowledge and skills of expert users, i.e., paleographers. An AI-based system can be designed, highlighting potential symbols to be validated and enriched by the experts, whose decisions are used to improve the detection performance. This paper shows how this task can benefit from feature auto-encoding, showing how detection performance improves with respect to trivial template matching.