Research On Building Facade Completion Network

The complex urban terrain environment,with many distractions such as vegetation and pedestrian and vehicular traffic,inevitably obscures the building facade,and these obscurations can have an impact on subsequent applications such as facade semantic segmentation,facade analysis,building structure restoration,and facade reconstruction.The problem of how to recover the occluded part of the facade image is crucial,which still faces the following three challenges at this stage:(1)building facade complementation lacks the support of a large number of annotated data sets,which makes it difficult for many classical deep learning-based algorithms to retrain a high-precision image complementation model;(2)unlike images of human faces and natural textures,building facade images are man-made scenes,which have structural features,dimensional features,regular features,etc.in addition to texture.(2)Unlike images such as human faces and natural textures,building facade images are man-made scenes with structural,dimensional,and regular features,and it is difficult for existing image completion algorithms to maintain the consistency between local and global textures and structures of building facades,and their completion results often do not conform to architectural a priori;(3)Although the current deep learning-based image completion framework introduces a contextual attention mechanism to focus on the long-distance dependence between the completion region and global features,it does not have a better This often leads to distortion of the edge structure of the target region,blurring of the texture,etc.1)Aiming at the problem of the complex structure of a building facade that is difficult to be repaired,this thesis proposes a building facade completion network based on semantic constraints.The algorithm uses the building facade parsing module to synthesize the semantic structure information of the building facade.First restoration of the semantic structure map and finally guides the building facade complementation network to obtain the final building facade image according to the complete semantic structure map.This effectively improves the model’s learning of complex building facade semantic information,thus improving the quality of building facade complementation.2)Aiming at the problem of it is difficult to train existing image complementation network models for building facade data with few samples,this thesis designs a dynamic convolution-based completion network with powerful feature learning capability,which can process the features of image data with global perception field in few sample data sets,ensuring the correlation between complementation regional features and global features,and improving the consistency and integrity of building facade complementation results in terms of global structure.3)Aiming at the problem of the missing areas of the building facade is difficult to reasonably repair,this thesis introduces a spatial attention branch in the dynamic convolution-based completion network,which enhances the feature representation of the part of the complementary area by strengthening the features on the edge of the mask and weakening the background features,effectively strengthening the correlation between the complementary area of the building facade and the local structure of its edge,and improving the complementary results.