Recognition And Extraction Of Color Steel Building Information Based On Deep Learning

Due to its advantages of low cost and quick and easy construction,temporary color steel plate buildings are constructed on a large scale in urban villages,urban fringe and various industrial parks and industrial parks.Its abundant existence,on the one hand,reflects the process of urban development;on the other hand,its material is easy to cause fire,making it a key monitoring target.Rapid and accurate extraction of building information of color steel plate is the basis of related research.In high resolution images,the differences of different types of color steel buildings made of different materials such as cause its on the edge of the spectrum,texture and so on have bigger difference,increasing the difficulty of extraction,recognition and extraction using conventional extraction method is difficult to achieve accurate,scrawled color steel buildings,it may be mistaken to mention and leakage has yet to have specifically targeted at color steel building information extraction method,based on the study of the extraction method of related buildings,in view of the excellent performance deep learning,In this paper,three deep learning improved extraction methods based on typical coded-decoding neural network,low level feature retention and channel attention,and sc SE and pyramid pooling are proposed to achieve accurate and fast extraction of color steel plate buildings.The main work is as follows:(1)Construction and processing of color steel building data set.At present,there is no public high-resolution remote sensing image data set of color steel plate buildings.Therefore,this study needs to complete relevant processing and production of deep learning data set of color steel plate.In this paper,domestic GF-2 satellite images were adopted and Arc GIS10.2 image processing software was used to complete the vectorization work of color steel plate buildings.The related processing was carried out by Python language to construct the deep learning sample database of color steel plate buildings,providing data basis for subsequent experiments.(2)Based on the classical coded-decoding segmentation network,the color steel plate building information is extracted.Two classical image segmentation networks,Segnet and U-net,were used to extract the color steel plate buildings.The network structure characteristics of typical codec image semantic segmentation were analyzed.The network structure was created and the extraction experiment was completed by using the self-made color steel plate sample data set.Combined with the task of color steel building extraction,this paper analyzes the effect of network extraction from qualitative and quantitative perspectives,evaluates the advantages and disadvantages of the network,and selects the neural network model which is more suitable for the extraction of color steel building from remote sensing images.(3)Propose an improved U-net network model reserved for low level features.Aimed at the limitation of the existing network,at the same time,to enhance the extraction accuracy of color steel buildings,inspired by YOLOv3 network,in this paper,the basic framework of the U-net network code path,and convolution is characterized by copy,the original features as input to the next level network,the characteristics of the replication by "sampling on pooling + convolution +" the combination of form and decoding path integration characteristics of "transition layer",after the convolution operation,through the attention module and decoding realize jumping connection path corresponding level characteristics,build low-level features retaining structure,which reduces the loss of coding the low-level features,At the same time,it strengthens the ability of the network to pay attention to the detail information and realizes the more accurate extraction of the color steel plate building.(4)In order to further optimize the extraction effect of color steel building,based on the U-net network framework,in the coding path of the network,based on the spatial pyramid pooling method,the asynchronous long pooling window is used to generate the feature information of the same size as the coding path level,and the jump connection is made.At the same time,the sc SE attention module was added after the model convolution operation to further improve the segmentation ability and extraction accuracy of the color steel plate building of the network.Finally,the two kinds of color steel building extraction networks proposed in this paper were compared with the typical semantic segmentation neural network,and the extraction results were quantitatively analyzed through the constructed quantitative evaluation index system,so as to verify the feasibility of applying deep learning to the extraction of color steel building and select the network model more suitable for the extraction of color steel building.Experimental results show that using different color depth of learning methods applied to steel building extraction has good extraction effect,verified the deep learning method to extract the feasibility of color steel construction,at the same time,the low-level features to retain and attention mechanism in the code path,pyramid pooling embedded network can ascend to a certain extent effect and precision of the model to extract color steel construction,this paper constructs two kinds of improved U-net model in accuracy,frequency power occurring simultaneously,Dice similarity coefficient,the kappa coefficient classification accuracy on indicators are better than that of the typical codec neural network segmentation,It is proved that the model proposed in this paper has good accuracy and robustness for the extraction of colored steel plate buildings.This study can better complete the task of color steel building extraction,color steel building extraction and the corresponding research has a certain reference value.