Research On Vectorization Of Scanned Map Line Elements Based On Deep Learning

The map contains rich geographic element information,which is an important data source in the geographic information system.However,all kinds of geographic element information of the scanned map exist in raster format data,and the geographic data in vector format is needed in the geographic information system,and the map data in vector format is easier to save than the paper scanned map.Therefore,it is of great significance to vectorize scanned raster map data.Due to the complexity of map image and the complexity of map line elements,traditional algorithms can not be well implemented in map image processing.With the maturity of image processing,deep learning,big data and other related technologies,compared with traditional algorithms,it is more convenient and fast to use these computer related technologies to realize automatic vectorization of map line elements.In particular,convolutional neural network in deep learning has a good application in image processing,which can automatically learn and acquire features of different scales of images.Therefore,this article intends to use convolutional neural networks to achieve the acquisition of map line feature features,identify line features,and classify them based on pixel level.Combined with relevant vectorization algorithms,a convolutional neural network-based automatic vectorization method for line features is proposed.The main research contents and methods are as follows:(1)Recognition of Map Line Features Based on GASRNet Network.In view of the problems of the original U-Net in the extraction of map line elements,such as line element detail loss and edge blur,the GASRNet model is proposed.Compared with U-Net model,it adds two new modules to realize the increase of the weight of map line elements and the acquisition of the features of map line elements of different scales in map images.The experimental results show that compared with U-Net model,the extracted map line elements are more complete,less affected by text interference,and can capture more map line element information.In the raster map data set,the accuracy of the improved network online feature extraction has been improved by 6%～8%,and the anti-noise ability has been improved by 13%～17%.In the scanned map data set,the accuracy of the improved online feature extraction has been improved by 7%～9%.Therefore,the improved network model proposed in this paper has certain practicability.(2)Scan Map Line Feature Classification Based on transfer learning.A simulated map dataset is generated by randomly adding text and background graphics to simulate the style of scanned map centerline features.In transfer learning,the source domain data is pre trained with the simulated map data using the Mobile Net,Res Net50,Inception V3,Mobile Net V2,and Xeption models to obtain the characteristics of map line elements,and then the model weights obtained are migrated to the scanning map training model to complete transfer learning.Experimental results show that using transfer learning can further improve the accuracy of the model for line feature segmentation.In the final pixel level classification of various types of line elements,the network model formed by the combination of Res Net50 and Segnet performs best.Compared to non transfer learning,the overall accuracy of transfer learning is improved by 1%～10%,the PA accuracy of various types of line element classification is improved by 7%～25%,and the IOU accuracy is improved by 11%～39%.(3)Automatic vectorization of scanned map line features.Based on vectorization algorithm,convert grid type line feature data into SVG format vectorized data for storage.In the experiment,the vectorization method of scanned map line features based on deep learning in this paper was compared with traditional vectorization methods.The method proposed in this paper has good vectorization effect and performance,and has a high degree of automation.