Street View Image Classification Based On Active Learning

With the continuous development of the times,the increasing urban population has led to the scarcity of urban land resources.The rational use of land resources has become an indispensable part of the process of urbanization,so it is crucial to scientifically classify urban functional zone.The classification of zone based on the types of buildings in the zone is the main method of classification of urban functional zone.Traditional building images mainly come from remote sensing images.In recent years,with the development of electronic maps,such as Google Map and Baidu Map,the acquisition of street view images in cities has become relatively simple,and buildings are the most important part of street view images.Therefore,the urban functional zone category of the street view image can be judged by the type of building,that is,the purpose of urban functional zone classification can be achieved by street view image classification.Compared with remote sensing images,applying street view images to urban functional zone classification has two main advantages: first,street view images are easier to obtain;second,street view images have more external features of buildings.At present,the biggest problem of street view image classification is the lack of semantic labels.If street view image classification is through supervised learning,a large amount of training data is needed to support the training process of the classification model.Manually annotation will waste a lot of time and economic costs.In response to this problem,this paper proposes to apply active learning methods to the street view image classification.The main research contents of this paper are as follows:(1)Based on street view images,we made a double label dataset named BEAUTY(Building d Etction And Urban func Tional zone portra Ying)1.0 and determine the classification model for the active learning process.First,through observation and research on the BIG?GSV,it is found that there are some invalid samples,illegible samples and samples with ambiguity in the BIG?GSV.For these three problem samples,corresponding solutions are proposed.Then,a classification standard for urban functional zone was formulated,and the BEAUTY1.0 was created according to this standard,and street view images were divided into 4 urban functional zone categories.Finally,Res Net50 and Res Net101 are used for training and testing on the BEAUTY1.0,and the confusion matrix between the 4 categories is given.The experimental results verify the scientificity of the BEAUTY1.0 and can be used to urban functional zone classification.And through the test results of two classification models,Res Net50 is determined as the classification model of the active learning process.(2)A street view image classification method combining active learning algorithm and Mixup is proposed.Firstly,Mixup is performed on a small number of labeled samples.The enhanced samples are used for model training,and the trained model is used for testing of unlabeled samples.Then according to the active learning selection strategy and the model test results,select the samples that are conducive to model training.Manually annotation and continue to be used for model training.The experimental results show that the street view image classification method combining active learning and Mixup has obvious advantages compared with the traditional active learning method and random active learning method,and the confusion matrix between4 categories is given.Then,by analyzing the difference between the traditional active learning method and the street view images that need to be labeled during the active learning process,it is verified that the method in this paper can be effectively applied to the street view image classification problem.(3)A street view image classification method combining active learning and Fixmatch is proposed.Firstly,selecting samples that are less confident in the classification model through the Fixmatch training process,and then according to active learning selection strategies to further select samples that are conducive to model training and manually annotate them,and continue to use them for model training.Considering that the training method of Fixmatch is different from supervised learning,this paper proposes two combination methods of "full combination" and "uniform combination".The main difference between the two methods is the setting of each active learning iteration process epoch.The experimental results show that the street view image classification method combining active learning and Fixmatch has obvious advantages over traditional active learning methods and supervised learning methods,and a confusion matrix between 4 categories is given.Then,by analyzing the difference between the traditional active learning method and the street view images that need to be manually labeled during the active learning process,it is verified that the method in this paper can be effectively applied to the street view image classification problem.