Research On Furniture BIM Texture Restoration Based On Convolutional Neural Network

In recent years,the process of industrial digital transformation has been accelerating,and building digitalization has become an inevitable trend of transformation and upgrading of the construction industry.As one of the most effective technologies to realize building digitalization,BIM technology can provide digital expression for the physical and functional characteristics of building facilities,and provide reliable shared information resources for various decisions in the whole life cycle of the building.Therefore,the authenticity and integrity of the BIM model for the description of building facilities is an important basis for determining whether it is available.Among them,the texture of BIM model,as an important non-parametric attribute to describe the authenticity,plays an important role in the later application of model design and creation,three-dimensional effect rendering,BIM visualization management and operation and maintenance.At present,BIM models are mainly derived from modeling software and various cloud platforms,and the model texture provided by them is too single,so that the model visualization effect is difficult to meet the authenticity requirements.At the same time,there are a large number of real component images in the network,which contain rich component textures.Therefore,how to retrieve the most similar real component image through the BIM model and obtain its texture map has become the focus and difficulty of BIM texture restoration research.Considering the wide application of convolutional neural networks in the field of image cross-domain retrieval,this thesis takes the furniture BIM model as the research object.By optimizing the image retrieval algorithm,the real furniture image which is most similar to the furniture BIM model contour is matched,and then its texture map is obtained,so as to realize the restoration of furniture BIM texture.The main research contents are as follows:(1)Texture reduction method of furniture BIM based on VGG network.Firstly,in order to obtain the backbone network with the best retrieval performance,the classical convolutional neural network is selected to train the network retrieval of real furniture image data sets.The experimental results show that VGGNet16 has a better retrieval effect,and it is used to restore the BIM texture of the furniture.Then,experiments are carried out to fuse the shallow and high-level semantic features of the network.The experimental results show that feature fusion and shallow network semantic features can better depict the features of furniture images.(2)Furniture BIM texture restoration method based on multi-scale feature fusion.Combining with the feature fusion strategy,the shallow and high-level semantic features of VGGNet16 and the output features of the first fully connected layer are fused,and the second fully connected layer is removed.The fused network output features are used for furniture image retrieval.The method retains the semantic features of the network with different depths regardless of the weight.At the same time,removing the fully connected layer reduces the loss of network semantic features in computation.Experimental results show that the retrieval accuracy is improved by 1.46% compared with VGGNet16.The retrieved furniture image can be used to restore the BIM texture of furniture,which proves the effectiveness and feasibility of the method.(3)Texture restoration method of furniture BIM based on CBAM and multifeature fusion.Combined with the attention mechanism,the weighted fusion is carried out in the feature fusion link,and at the same time,the output features of the fully connected layer are fused with the HOG features to form new features for furniture image retrieval.The weighted fusion method enables the network to depict the prominent features of the image while retaining the original features,and at the same time,the fusion of HOG features can further fully depict the image features.Experimental results show that compared with VGGNet16,the network retrieval accuracy based on CBAM and multi-feature weighted fusion is improved by 1.8%,and the retrieval m AP value of network semantic feature and HOG feature tandem fusion is the best.Texture mapping is extracted from the retrieved real furniture image,and the restoration of furniture BIM texture is realized more realistically.Figure [58] table [15] reference [50]...