Intelligent Recognition Of Illegal Buildings Based On Deep Learning

Buildings are an important part of urban basic geographic information.If they are not supervised and planned,they will not only seriously erode urban public resources,endanger public safety,and affect the external appearance of the city,but also detrimental to urban management and development.Manual judgment is not only inefficient and wastes manpower,material and financial resources,and traditional extraction methods also mainly rely on artificially designed features such as length,edges,texture,shadow,spectrum,etc.such empirically designed features are generally only applicable to specific data,and the accuracy of the data is low,the generalization ability of the model is also not good.Therefore,it is necessary to study an automatic,accurate and fast extraction method of illegal buildings.In recent years,with the continuous development of hardware and deep learning technology,major breakthroughs have been made in image processing fields such as object detection and semantic segmentation.Many scholars have gradually used deep learning in building extraction and have achieved certain results.Based on the data from two different data sources of urban monitoring and remote sensing,this paper studies the intelligent identification methods for illegal buildings in the above two scenarios by improving and optimizing the existing deep learning methods.The main work of this paper is as follows:(1)Data set construction and data enhancement method research,detailed introduction of the construction process of urban monitoring data,remote sensing image data and data enhancement methods.The data enhancement method is used to increase the diversity of samples and make the generalization ability of the model.Stronger.(2)Analyze the illegal buildings in the urban monitoring data,and find that the illegal buildings are obviously different from other buildings in certain characteristics(such as the building is under construction,just completed,very dilapidated).Then,we improve on the basis of Mask-RCNN and study a method for extracting illegal buildings.In order to further improve the accuracy of the target detection frame,this method uses K-Means to cluster the monitoring data set,and applies the anchor point frame obtained by the clustering to the regional proposal network,then improve the semantic segmentation branch,use different convolution kernels to convolve the features processed by Ro I Align in parallel and fuse the features to improve the accuracy of the mask,and finally use transfer learning to fine-tune our model based on the pre-trained model of the coco data set.(3)In view of the richer detailed information of remote sensing images and the difficulty in distinguishing features,a multi-level residual network RSDANet with attention mechanism is studied to extract buildings,then overlaying the extraction results with the basic farmland protection area to determine whether the building is illegal.The network is built using the classic Encoder-Decoder structure.In the Encoder,a two-layer residual structure is designed,which can make full use of local features and the surrounding context information,then use a bidirectional feature pyramid structure to fuse multi-scale features,finally,the spatial attention,channel attention and weight adaptive methods are combined to further extract the interrelationships between objects in the high-level feature map.In the decoder,skip connections are used to fuse high-level features and low-level features step by step,making full use of spatial information to improve the accuracy of the prediction mask.In addition,new branches of boundary prediction are added to the network,and the accuracy of the model is further improved through the constraints and influence of boundary information.