Research On Building Inspection Based On Deep Learning

Building detection in remote sensing is significant for urban land planning,environment perception,and illegal buildings monitoring.Because remote sensing images have detailed and landform background information,traditional algorithms suffer from problems of missing building edges and low detection accuracy.To this end,it is imperative to develop new algorithms for building detections in remote sensing.With the advent of the big data era,building detection methods based on deep CNN(convolutional neural networks)have emerged rapidly.From the summary and analysis of building detection methods at home and abroad,this paper proposes new CNN based building detection method and achieves good results.The main work is as following:(1)Aiming at solving shape variations in different regions,a building detection method based on dense convolution is proposed.Traditional building algorithms are difficult to extract discriminative features from diversified buildings.This paper adding an internal dense convolution module to achieve the purpose of enhancing the feature extraction ability.After that,from the perspective of enhancing information flowing in the model,a dense short connection module is proposed to filter detailed and semantic information.This method improves the ability of extracting discriminative features,enriches the efficiency of information flowing in the network,and improves the detection accuracy.(2)Aiming at focusing on different types and layout between large and small buildings,a building detection method based on dual segmentation fusion model is proposed.This method applies the idea of hierarchical classification to divide buildings into large and small buildings and detect them separately.To adapt to urban cities,the large building segmentation model based on U-NET adds RESNET to get strong feature extraction capabilities and ASPP to get larger receptive fields.In addition,this paper fuses the shape and background information in feature fusion stage.In order to alleviate the imbalance problem for small building,the small building segmentation model adds a hybrid loss strategy to focus on small buildings.After fusing the prediction of the large and small segmentation model,the method integrates effective region information from the dual segmentation model,and further improves the detection performance.(3)Aiming at dealing with scale differences between large and small buildings,a building detection method based on the fusion of pixel and region segmentation models is proposed.In order to pay more attention for small buildings,this method introduces Mask R-CNN for small objection detection.And the pixel segmentation model also adds the SWA strategy to mitigating the imbalance problem of large and small buildings.On this basis,our method uses decision fusion to intercept effective detection intervals to improve the recall rate of building detection results and enhance the robustness of the model.