Automated Extraction And Spatial Reconstruction Of Scanned Map Annotations Based On Deep Learning

Map annotations are an important part of maps,and extracting annotation information from scanned maps helps to further query and analyze map information and fully exploit the value of scanned maps.Due to the diverse characters,complex lines and changing directions,the automatic extraction of Chinese scanned map annotations is very difficult and less researched at present.The rapid development of digital image processing technologies such as deep learning provides the possibility of automatic extraction of map annotations.In this paper,we propose a deep learningbased automatic extraction method for Chinese scanned map annotations,including three modules of scanned map annotation detection,annotation recognition and spatial reconstruction.The main research contents and conclusions are as follows:(1)Scanning map annotation detection based on improved Advanced EAST.In order to extract deeper image information,the Advanced EAST feature extraction network is replaced by Res Net50,and its network structure is adapted to the whole method.The F1-Score value of the final note detection model is 0.79～0.89,which is more than 0.1 higher than that of other similar network models.(2)Scanned map annotation recognition based on migration learning.In this paper,we propose a migration learning based notation recognition model.The note recognition model based on CRNN is subjected to two migration learning based on model sharing.The first migration learning takes the 3.6 million Chinese document dataset as the training set,which mainly makes the CNN and RNN networks learn as many kinds of characters as possible,and the second migration learning takes the CAPTCHA dataset as the training set,based on the first migration learning,the same CNN and RNN networks are learned,mainly to enhance the resistance of the model to line elements.The final obtained F1-Score values of the notation recognition model are0.89～0.93,which are more than 0.09 higher than the F1-Score values of other similar network models respectively.By superimposing with the notation detection model,the overall accuracy of extracted notations Precision,Recall,and F1-Score achieved more than 0.87,0.88,and 0.8,respectively,with better noise immunity performance.(3)Scanning map annotation classification and spatial reconstruction.A CBAM-UNet++ based annotation segmentation model is proposed using semantic segmentation and cluster fusion,which is based on UNet++ and incorporates CBAM modules into each convolutional stage.The annotation classification results and the annotation extraction results are projected to reconstruct the scanned annotation space location.The results show that the final accuracy of the model reaches 0.93,which is more than 0.05 improvement compared with other network models of the same type.The spatial reconstruction effect has a good agreement with the actual location.