Multi-class Wall Recognition in Complex Architectural Floor Plan Images Using a Convolutional Network.

Architectural floor plan is an essential document to share the building information among designers, and engineers. Automatic floor plan image analysis is useful to extract various information from the floor plan. Wall segmentation is an important step in floor plan image analysis. However, few research works have been conducted for automatic wall recognition in an architectural floor plan. In this paper, a convolution neural network, namely WallNet, is proposed to recognize the multi-class walls. The WallNet consists of an encoder and a decoder. The encoder captures low-level features, as well as multiscale contextual information. Based on these extracted feature maps, the walls are detected. The proposed network is applied to recognize five different classes of walls: solid-wall, dot-wall, diagonal-wall, hollow-wall and gray-wall. The experimental results show that the proposed architecture can obtain a mean average precision of 72%, which is superior compared to the state-of-the-art techniques.