Research On Classification And Information Extraction Of Ubiquitous Charts

In recent years,with the rapid development of Internet and intelligent mobile terminal technology,the breadth and depth of ubiquitous network also increase,which is accompanied by the explosive growth of ubiquitous information.The data format of the ubiquitous chart is raster image,so the information extraction of the ubiquitous chart can be accomplished by relying on the related technology of computer vision.Because different types of charts display data through different graphs,the basis of intelligent interpretation of ubiquitous chart information is automatic classification of ubiquitous chart,and then different information extraction methods are adopted according to the type of chart label.However,because the ubiquitous information has the characteristics of discretization,high noise,strong fuzziness,unstructured,it is difficult to achieve good results in the classification and information extraction method based on single feature at present.Therefore,according to the characteristics of ubiquitous information,it is of great practical significance to study the automatic classification and information extraction methods suitable for ubiquitous graphs.In our approach,an in-depth study is carried out on the classification and information extraction methods of ubiquitous charts.By analyzing the problems existing in the current classification and information extraction methods,four common kinds of ubiquitous charts classification model is proposed,and the visual characteristics of ubiquitous charts are summarized.Finally,the information extraction methods of ubiquitous charts are explored by taking bar charts as an example.(1)An automatic classification model of ubiquitous statistical charts based on transfer learning is established,and the classification model is improved by parameter optimization and model structure optimization.(2)To solve the insufficient generalization performance problem of ubiquitous chart automatic classification model,based on the theory of map symbol visual variables,the visual features of ubiquitous charts are quantitatively described.A data augmentation method considering the visual features of ubiquitous statistical charts is proposed and the training data set of ubiquitous charts is augmented by using this method.(3)In the aspect of chart information recognition and extraction,a method of ubiquitous chart information recognition and extraction based on Paddle OCR and color characteristics is proposed.The principle of various information extraction algorithms is illustrated by taking column chart as an example,and the feasibility of the algorithm is verified by experiments.