Research On Recognition Of Physical Circuit Schematics At Secondary School Based On YOLO

With the emergence of AI 2.0 and Big Data,intelligent tutoring system has developed sig-nificantly,simulating the teaching tasks of educators and designing appropriate teaching strategies to help students to acquire new knowledge and solve problems.Automatic solv-ing as an important part in intelligent tutoring system,provides key technology support for primary and secondary education.Automatic solving for physical circuit problems as a sub-problem of automatic solving is facing difficulties such as natural language understanding,physical circuit schematic recognition and analysis.This thesis presents a series of methods for automatically recognising circuit elements and converting circuit schematics into circuit graphs by YOLO(You Only Look Once).The recognition of physical circuit schematics,as the basis for automatic solving of physical circuit problems,includes localization and cate-gorization of all elements in physical circuit schematics,and the circuit structure recognition without the identified circuit elements,which can not only make the structure analysis of cir-cuit schematics more clearly,but also facilitated the relation extraction from physical circuit schematics.A novel approach is to model physical circuit element recognition as a multi-object de-tection problem by using YOLO to localise and classify each element.In this thesis,the deep object detection algorithm is used to solve the problem.Object detection is a com-mon computer vision problem,which involves the recognition and positioning of objects and the recognition of certain categories in images.The object can be positioned in a variety of ways,including creating a border around the object or marking each pixel that contains the object in the image.For example,traditional machine learning methods combined with sliding window method are considered,which requires training a machine learning classifier in advance.Although this method makes the recognition accuracy higher,it is necessary to train the machine learning model first and the recognition speed is too slow.Thus,deep object detection algorithms are considered to solve the problem.After comparing multiple methods,YOLO is decided to be used in the research.The thesis also explains how to use YOLO and the reasons to use YOLO.Another novel way is to use improved LSD(Linear Segment Detection)which is suitable to recognize physical circuit structure and often appears in computer vision to detect lines.LSD is the key to analyse physical circuit schematics because circuit schematics can be trans?formed to structured data of graph theory according to the low-level features extracted from original circuit schematics,but there are still some defects of LSD need to be amended to better present the structure of circuit schematics.To amend the defects of the original LSD and convert schematics to graphs,a series of algorithms are presented.This thesis presents a series of algorithms for recognizing physical circuit schematics ap-peared in the circuit problems at secondary school.The problem for recognizing physical circuit schematics can be divided into two parts which are physical circuit element recogni-tion,and circuit structure recognition.The challenges include not only the collection of la-belled data and training of YOLO but also the transformation from physical circuit schemat-ics to circuit graphs.Besides,circuit graph analysis is also a tough challenge.To deal with the challenges,YOLO,a deep learning object detection method,is used to localise and clas-sify 13 kinds of elements in the physical circuit schematics,and linear segment detection algorithm and the proposed algorithms are used to convert physical circuit schematics into circuit graphs.The thesis also discusses the extraction method for some circuit relations.The circuit rela-tions include the relations between current,voltage and resistance,which can be analyzed and extracted through some algorithms of graph theory.The extraction of the physical cir-cuit relations is the most important step in solving the physical circuit problem,but there are many kinds of relations in physical circuit,and the focus of the thesis is mainly on the recog-nition of the physical circuit schematics,and only the extraction of the KCL(Kirchhoff s Current Law)and KVL(Kirchhoff's Voltage Law)physical equations is discussed.Thus,there is no much discussion about circuit relation extraction in this thesis.However,in the future work,extraction of physical circuit relations will be further expanded based on the recognition methods of physical circuit elements and physical circuit structures proposed in this thesis,not limited to extracting KCL and KVL equations.The extraction of the physical circuit relations is the basis for the automatic solving of physical circuit problems,and the correctness of the solution basically depends on the extracted physical circuit relations to a large extent.Thus,in the later research,the emphasis of the research will be placed upon the physical circuit relation extraction.In the thesis,the independent nodes and loops are identified according to the positions and types of physical circuit elements and the structure of physical circuit schematics.Therefore,the extraction of physical circuit relations depends much on the recognition of physical circuit schematics.The experiment of the research is performed on the physical circuit schematics collected from the commonly-used secondary school textbooks,entrance examination papers and tutorial books.The circuit schematics collected in the data set are highly representative,because all the data are selected from the books or the papers without filtering,and the selected books or the papers are used by more than 20 million Chinese students each year.During the experi-ment,the accuracy of YOLO in the recognition of physical circuit elements was evaluated,and the performance of YOLO and Faster-RCNN on some circuit elements was compared.In general,YOLO has higher performance on physical circuit element recognition than Faster-RCNN.In addition,the mesh search method and the circuit analysis algorithm proposed in the thesis have also been compared on the accuracy of different methods for the extraction of KCL and KVL equations.In conclusion,the thesis mainly introduces the recognition of physical circuit elements and physical circuit structure,and the structure of physical circuit schematics can be saved in the form of a graph through a series of algorithms,so that the algorithms in graph theory can be used to analyze the physical circuit structure to extract the circuit relations.Recognition of physical circuit schematics is one of the foundations for the automatic solving of physical circuit problems,which is helpful for the development of intelligent tutoring systems.