Detecting Changes In Buildings From Aerial Images Based On Deep Learning And Matching Region Reconstruction

Building changes detection is a technique that analyzes two images with the same area acquired in different time periods to obtain building changes information,and plays an important role in urban planning,land and resources management,and geographic information updating.With the rapid development of aerial photography technology,building changes detection based on aerial images has become easy to implement.However,the existing methods suffer from problems such as non-building interferences when detecting the whole image,the limited use of building features,and the unbalanced number of positive and negative training samples.In addition,the false alarm rate of the existing method is increasing when buildings in two aerial images have different angles.To this end,based on the aerial images provided by the Dalian Geotechnical Engineering and Mapping Institute CO.,LTD.,this thesis has carried out the following three aspects of work:(1)A non-building interferences removal method is proposed based on U-Net deep segmentation network.First,through color space conversion,the chromatic aberration between buildings in the image to be detected is weakened,and the influence of illumination difference is reduced.Then,the accuracy of the buildings extraction is improved with a small amount of data via data augmentation and image post-processing.Finally,according to the extraction results,the image logic operation is performed to obtain a mask image of the non-building interferences removed for building changes detection.The experimental results using the existing change detection method show that the proposed method improves the accuracy of detection and reduces the amount of calculations.(2)An improved change detection method based on unsupervised deep network PCANet feature extraction and SVM classification is proposed,and building changes detection is performed on the non-building interferences removal image obtained by this thesis.The existing method only utilizes the spectral features of the image,while the texture features can better characterize the spatial structure of the buildings.To this end,this thesis combines texture features and spectral features to enhance the ability to express changes information in buildings.Then,by oversampling the pixels with smaller number,the number of positive and negative training samples is balanced,and the effectiveness of the deep-level features of changes extracted by PCANet is further improved.The experimental results show that the improved method increases the accuracy of change detection compared with the pre-improvement method.(3)A two-round matching region reconstruction and comparison method is proposed for building changes detection with different aerial angles.First,a single building pair is extracted from the ground truth buildings segmentation,and the matching points are extracted using the deep network.Then,using the boundary matching points to perform the first round matching region reconstruction and comparison,the preliminary determination of the building changes is realized.When the aerial angle is small,the first round of determination is valid.When the aerial angle is large,continue to apply the structural similarity-based search strategy,reconstruct the most similar single building pairs,and perform the second round of matching region reconstruction and comparison.The experimental results show that the method can reduce the false alarm rate of changes detection and improve the accuracy.