Method And Implementation Of Architectural Floor Plan Structure Recognition Based On Multi-task Model

With the rapid development of computer technology,there are many demands and applications that need to obtain the internal structure of buildings,such as Internet home decoration,online real estate transaction,Simulation and simulation of indoor applications and so on.Architectural floor plan is the carrier of expressing building information,which generally includes walls,doors,windows,rooms,furnitures and some annotation information such as text.The main task of architectural floor plan structure recognition is to obtain the geometric information of walls,doors and windows and rooms formed by their connection.The difficulty of architectural floor plan structure recognition is that there are great differences between architectural floor plan,and the sources of differences include drawing styles and different expressions of the same kind of objects.To solve the problem of architectural floor plan structure recognition,this thesis first gives a vector-graphics representation based on the existing vector-graphics representation,which hierarchically expresses the structural information of architectural floor plan,and gives the implementation details of data annotation algorithm.Then this thesis designs a two-stage architectural structure recognition method,which includes a architectural structure element recognition network based on multi-task and a architectural structure reconstruction algorithm based on structural constraint.The architectural structure element recognition network is a multi task network model combining line segment detection and semantic segmentation.Line segment detection can predict the line segments representing walls,doors and windows.Semantic segmentation can predict the segmentation information of walls,doors,windows and rooms.The architectural structure reconstruction algorithm based on structural constraints constructs the hierarchical relationship between line segments representing walls,doors,windows and room instances,that is,the room contour formed by line segments representing walls,doors and windows.The method is tested on two different datasets,and the experimental results fully illustrate the effectiveness of this algorithm and achieve good results.In addition,this thesis designs and implements a architectural floor plan recognition system based on the designed architectural structure recognition method.This thesis introduces the system in detail from three aspects: the overall design,system implementation and system application.