Research On Task-Driven Visual Computational Models For Mathematical Educational Resources

The deep fusion of Internet,intelligent technologies and education enriches not only the methodologies of education,but also the learning behaviors.The development of intelligent education is one of the important strategies for China’s education reform,and the utilization of intelligent technologies can help to promote the reform of teaching methods.The wide application of various educational platforms and software have accumulated a large number of digital educational resources,which provide a variety of options for users’ self-learning.However,these resources vary and are not ensured in quality,seriously affecting the efficiency and effectiveness of users’ exploration.This brings an urgent need for multi-faceted processing of the digital educational resources,and it is important to study the processing methods and techniques applicable to digital education resources.Relying on the Guangdong Provincial Engineering and Technology Research Center for Mathematical Education Software,this dissertation focuses mainly on three important tasks in constructing high-quality educational resources,i.e.visualization of the resources,visual processing and editing of the contents and user interface layout organization.The main work and contributions include the following four aspects.To deal with the visualization task of mathematical education resources,a visualization model based on the traversal of interface elements is proposed.First,the structured description of mathematical educational resources is defined,then the interactive elements in mathematical educational resources are extracted,and the dependencies between the elements are defined and analyzed.In order to ensure the visualization results consistent with the resource as much as possible,the rules of interaction are set,and based on these rules the algorithm for interaction sequence generation is designed.Making use of the idea of graphical user interface testing,the generated interaction sequence is traversed automatically by the written scripts.The traversal is recorded into video that presents the knowledge in the resource,realizing the purpose of dynamic visualization.The proposed model has been applied to Super Sketchpad,a dynamic mathematics educational system,and its validity has been verified by user studies.In terms of the visual processing and editing tasks of educational resources,after analyzing the visual editing problems in the images of educational resources,taking advantages of some research results in deep learning,the dissertation studies the computational models of the visual contents of educational resources,and proposes the edge-preserving filtering models based on convolutional neural networks that effectively and efficiently imitate various filtering methods and address the problems of color shift and gradient inversion,as well as improve the filtering efficiency.Considering the color requirements for the images used in educational resources,the dissertation puts forward a colorizaiton model based on multi-cues manipulation.The whole process of local color transfer includes the steps of color clustering,inpaintingbased boundary correction and gradient-preserving optimization.The proposed model can not only achieve a visual satisfactory local color propagation results,but also well preserve the texture details.In view of the task of the interface layout organization of educational resources in mobile applications,an approach on user interface(UI)layout rearrangement based on the nodes-pairing model is proposed.To provide users with better experience,the interface should automatically adapt to a variety of sizes of display devices,display scales and resolutions,the UI components need to be arranged reasonably and their positions in the layout container should be decided.Building a hierarchy for the interface layout is the key to make the UI adaptive.Considering that the UI layout is constructed from the leaf nodes(basic components)to upper-level nodes level by level,the dissertation proposes a nodes-paring model and train the model with the machine learning algorithms with the UI component features abstracted and analyzed from a UI gallery.The trained model is then used to output the component layout with the input absolute data of a given UI.In the experiments on 3362 UIs of 800 open source applications,the average accuracy of the approach reaches over 90%.With the approach,the complexity of UI organization can be reduced effectively.In order to further improve the efficiency of resource organization,this dissertation proposes an improved clustering algorithm,the density-based clustering with maximum intra-cluster distance and adaptive Eps(DBCMIDAE),and further proposes a UI example matching model based on the DBCMIDAE.With the model,the UI examples are automatically extracted from UI search results.Comparing with the UI examples of six common functions of software such as "Sign in" in Apache,the experiments show that the results of the proposed model cover 83.33% of the examples summarized by Apache,and also provide some reference examples of different styles that are not included in Apache.It indicates that the proposed model can not only extract examples,but also find some new valuable examples to support the efficiency of UI organization.This dissertation explores some of the main problems that need to be solved in the construction of high-quality educational resources.Driven by the three tasks of visualization of the resources,visual processing and editing of the contents and user interface layout organization,this dissertation proposes the corresponding models and methods,and at the same time verifies the effectiveness of the work from objective data and subjective effects through the experiments.The work is expected to give an idea for the construction of high-quality educational resources.