Attention Enhanced Interactive Sketch Recognition And 3D Modeling Of Building Windows

As the skeleton content of urban environment,3D realistic and delicate building model is the key basic information of intelligent applications such as smart city construction,urban security and VR / ar.The window structure of the building facade is the main component of the building,which contains rich architectural geometric and semantic information.However,the existing research usually simplifies the window structure in the three-dimensional modeling of the window.The reconstructed window model often only has the boundary contour information,and lacks the restoration and description of the fine geometric structure of the window.This paper presents a reconstruction method of building facade window model based on interactive sketch,which can quickly and accurately reconstruct highprecision 3D window model.The existing building and facade reconstruction methods represented by City GML only use plane structure to express the window,without semantic definition and division of window frame,window sill and other structures.The reconstructed window model lacks type and fine structure information.In order to better define and express the typical window structure,this paper studies and implements the window model generation method based on grammar parameters.The types of windows with diversified styles and structures are summarized,and the generation and modeling processes of different types of windows are summarized.The generation grammar of windows is designed and implemented based on the General Markup Language XML.At the same time,a variety of parameters are defined to describe the fine geometric structure of windows.Procedural modeling requires users to manually edit grammar parameters,which is not easy to use.The sketching method can recognize 3D model types and parameters based on sketch images through intuitive and simple interaction,reliably restore regular object types and geometric structures,and is suitable for fine 3D reconstruction of windows.This paper studies and proposes to use the hand drawn sketch as the input.By identifying the predefined grammar parameters instead of the user editing process,the system automatically outputs the grammar types and parameter values required for modeling,and generates the final threedimensional model.In order to identify the window type corresponding to the target from the sketch and estimate the parameter value during modeling,two types of convolutional neural network models are studied and proposed in this paper.The classification network is designed and implemented based on Res Net framework.It classifies the input sketch and outputs the probability of the corresponding grammar of the sketch.In the design of parameter estimation network,attention mechanism is introduced to focus on the structural differences of windows.Based on the Res Net framework,a convolution attention module including both channel attention and spatial attention is integrated to improve the accuracy of parameter estimation.When training two kinds of network models,an automatic simulation generation method of non-photorealistic sketch image based on stroke rendering is realized,which replaces manual sketching to generate a large number of sketch data to train the model.The experimental results verify the effectiveness and feasibility of this method.When reconstructing the building facade window model,this method can quickly reconstruct the high-precision window model by simply sketching.The modeling efficiency is higher than that of Sketch Up and other modeling software.It only takes a few seconds to draw the sketch to generate the model.The performance of the two kinds of networks is evaluated.The top3 accuracy of the classification network for sketch recognition can reach 99%.The parameter estimation network with attention mechanism is also better than the current mainstream network in accuracy.Even if there is a gap between the drawn sketch and the target,the output model can be identified correctly,and the system has good robustness.