Research On Building Recognition From Remote Sensing Images Based On Multi-task Learning

Remote sensing image is a kind of image which can record the ategory attribute and distribution characteristics of ground objects according to a certain proportion,and can express rich surface information.The interpretation of remote sensing images has been widely used in land surveying and mapping,urban expansion,environmental detection,urban planning and construction,natural disaster emergency management and many other fields.In particular,building recognition in remote sensing image is the key link and research hotspot of remote sensing information acquisition.In this paper,a deep neural network based on multi task learning is proposed to explore the method of building recognition in remote sensing image through semantic segmentation and instance segmentation.The specific work is as follows:1.In view of the problems of the loss of detail information and the low accuracy of multi-scale building recognition in the traditional based semantic segmentation method in building area recognition of remote sensing image,a new building recognition method based on semantic segmentation is proposed in this paper: firstly,the whole model is divided into three parts: encoder,decoder and jump connection according to the "U" structure;then Xception module is used improves the model of encoder CNN module to improve the model performance;secondly,the improved Re Net model is used as the RNN module to reduce the computational overhead without affecting the model performance,and this structure is connected in series with the CNN module to extract image features;finally,the astrous spatial pyramid pooling is improved according to the characteristics of different sizes and shapes of buildings in remote sensing images in order to improve the recognition ability of the model for buildings of different sizes.Through experiments,the average overall accuracy of 97.78%is obtained on the WHU Building dataset data set,and 89.12% on the Inria data set,which is 0.05% and 4.27% higher than that of Seg Net,respectively.2.Firstly,aiming at the limitation of building area recognition based on semantic segmentation in building recognition algorithm of remote sensing image based on deep neural network,a neural network model based on multi-task learning is proposed.The multi-task network can complete the building area recognition based on semantics and individual building recognition based on instance segmentation.According to the task requirements,the data set is labeled with semantic segmentation tags,so that the data set has both semantic segmentation tags and instance segmentation tags.Secondly,it is difficult to find the correct balance between the sub tasks when minimizing the multitask loss function value in the multi-task deep neural network training process.An adaptive weight algorithm is proposed to automatically assign the weight to the sub task loss function,so that the multi-task can achieve the balance as much as possible in the training,and improve the generalization ability of the model.The experimental results show that the average overall accuracy of the multi-task network model is 98.20%,which is 0.49% and 0.22% higher than Seg Net and Deep Labv3+ based on VGG16,respectively.Finally,in view of the existing research on building recognition in remote sensing image,it is only based on semantic segmentation to complete the regional recognition of buildings in remote sensing image and case-based segmentation to complete the individual recognition of buildings in remote sensing image,but not according to some characteristics of buildings,it is proposed to label and make data sets according to the needs,and then use the method of building recognition based on semantic segmentation The neural network model of multi-task learning realizes the task of identifying school buildings.The average F-score was 61.69%.