Buildings Detection And Classification Based On Streetscape

With the rapid development of urbanization,the layout planning of buildings becomes particularly important for the development of cities.Remote sensing images were used in the early urban plan,but the information provided by remote sensing images was not detailed enough and mainly from top view.Compared with the remote sensing images,the street view images contain more scene information from different directions and can reflect more detailed features of buildings.Therefore,building detection and classification can effectively improve the efficiency and accuracy of urban regional evaluation and expand the evaluation scope.The research results will provide data support for city fine management and design.By researching the data sets containing street buildings,it can be found that the relative data set used to image classification is scared.The BIC?GSV dataset is a dataset with the detailed information.But two problems are obvious.One is that there are several objects in an image,but only a catagaory tag is given,so it can not accurately describe each kind of category information.The other is the wrong tag in some pictures which may be labeled inaccurately due to subjective reasons of the annotator.In addition,there is no suitable method for multi-label detection and classification of buildings.In view of the above problems,the research contents of this paper are as follows:(1)The calibrated multi-label BDISV1.0 street view building dataset was established based on the BIC?GSV dataset.First,the main buildings of the images in the BIC?GSV dataset were selected,the object building meeting the requirements was marked with boundary boxes,and its tag was determined under the guiding of the experts.Lable Img software was used to calibrate.(2)The training samples in the same category may vary greatly and the numbers of samples are not balanced for different catogories,so Precision(P),Recall(R)and F1 score(F1)can no longer be used to evaluate the classification effectiveness.Therefore,on the basis of VOC 2007 mean Average Precision(m AP)evaluation,the mean Average Recall(m AR)and mean Average F1 score(m AF1)were proposed to evaluate the experimental results in this paper.(3)The Single Shot Detector(SSD)method based on Convolutional Neural Network(CNN)is proposed to detect and classify the building with the multiboxes and tags.The experimental results show that the prosposed method can effectively detect and classify the object buildings for the established dataset comparing with the original data set with single tag.To some extent,it can avoid the confusion of multiple buildings with different class in an image,which is a relative serious problem in an image with single tag.In addition,the proportion threshold of target bounding box is decreased to solve the detection failure problem since the part of a building can be only detected because of shielding.On the other hand,the position information of the buildings can be ignored since the distance between the buildings on an image is small.Thus,the question is converted into multi-label classification,which greatly improves the classification effectiveness of the buildings.(4)A sample enhancement method based on ring-symmetric Gabor transformation was proposed.Although the data has been enhanced in the spatial domain in the adopted migration SSD network,the data is expanded using the ring-symmetric Gabor transformation to enhance the texture features of the buildings since the texture information of building with different category is obviously different.Experimental results show that the proposed method can improve the accuracy of building classification to some extent.