Floor Plan Semantic Segmentation Using Deep Learning with Boundary Attention Aggregated Mechanism

Floor plans play an essential role in the architecture design and construction. It serves as an important communication tool between engineers, architects and clients. Automatic identification of various design elements in a floor plan image can improve work efficiency and accuracy. Prior research has used image analysis and/or convolution neural network (CNN) to detect wall lines and segment rooms based on text annotations. However, text annotations may not be available and an efficient technique to automatically segment multiple elements of a floor plan is required. In this paper, a CNN-based technique is proposed to detect elements such as wall, door, and bedroom, and segment the floor plan by developing a room boundary attention aggregated mechanism. The room boundary prediction is performed simultaneously with the room type prediction. The attention mechanism makes use of the well-predicted room boundary feature to benefit the room type prediction. The analysis shows that a clear and well-shaped boundary appears in the attention model when the mechanism best improves the network performance. Experimental results show that the proposed technique can achieve a better performance than the state-of-the-art methods. The overall accuracy of the proposed technique for 9 categories classification at pixel-level is more than 92%.