Automatic Extraction Technology For Historical Observation Data Of Ionospheric

Ionospheric disturbances can have an impact on communication,broadcasting,positioning and navigation.In order to predict the changes of the ionosphere in time and reduce the impact on human life,the Chinese Institute of Radio Propagation carried out observation and research,and retained the vertical measurement report data images of ionospheric observation sites since the 40 s of last century,which are mainly divided into three types: printed data images,handwritten data images and dot matrix data images according to fonts,and the correct identification of these data is conducive to later data management and data mining.In this paper,for the historical observation image data in the specific field of ionospheric observation,this paper studies image preprocessing,text detection and text recognition algorithms for the purpose of improving the text recognition rate,and the main work is as follows:Firstly,the preprocessing methods of three ionospheric historical observation data images were studied.Aiming at the problem of missing border lines in the text image of printed data,a method of reconstructing the text frame line by horizontal projection and vertical projection is proposed,and the image is segmented on the basis of supplementing the text frame line.The contour detection method is proposed to extract the table area by the contour detection method of the handwriting data image,and the table area image is divided according to the divided line.For the dot matrix data image,a template matching method is proposed to match the segmentation sign characters in the image,so as to extract the text area to be recognized in the image and complete the preprocessing of the dot matrix data images.Secondly,the text detection algorithm based on deep learning is studied.The character distribution in the printed body and dot matrix data images is relatively regular,and the DBNET with Res Net50 as the backbone network is used to detect the text.Aiming at the problem of poor detection effect of handwriting data image in this network,DBNET with Mobile Net V3 as the backbone network is proposed to detect handwritten data images.By comparing the detection results of the commonly used CTPN and EAST text detection networks with the detection networks in this paper on ionospheric data images,the detection algorithm proposed in this paper has a good detection effect.Finally,aiming at the problem that the recognition accuracy of CRNN network is difficult to meet the project requirements,the recognition of ionospheric data images is realized by improving CRNN.To solve the problem of misidentification caused by character similarity,similar characters are counted,the dataset is expanded,and the model is trained.The recognition algorithm proposed in this paper makes the character recognition rate of printed data image reach more than 98%,the character recognition rate of handwritten data image reach more than 97%,and the character recognition rate of dot matrix data image reach more than 96%.On the basis of text detection and text recognition,in order to store the recognition results in tabular form,the RARE algorithm is used to realize the tabular recovery of ionospheric historical observation data.