Research On Circuit Components Recognition For Automatic Problem Solving

In the era of ICT in education and artificial intelligence,to develop the intelligent tutoring technology with the advantage of artificial intelligence becomes a hot research problem.As an important part of intelligent tutoring,automatic problem solving(APS)refers to the automatic understanding,reasoning and solution generation of problems by the machine.For a circuit problem,the problem understanding includes text understanding and diagram understanding.The diagram understanding for APS needs to extract the knowledge hidden in a circuit,based on the result of component recognition.However,there are still several challenges need to be solved in the component recognition for online learning application with massive users.This thesis employs the current popular deep learning methods and conducts two aspects of optimization to deal with the problems encountered in different application scenarios.Firstly,based on the centralized recognition application scenario of cloud services,the accuracy by applying universal deep learning algorithm can not meet the actual demand,which needs to be improved.Secondly,based on the distributed recognition scenario for mobile terminal,the current methods do not take into account lightweight deployment demand,such as lack of computing power and sufficient storage space.Therefore,in order to meet the actual deployment requirements of physical circuit component recognition,this paper attempts to introduce deep learning methods,respectively proposes two component recognition optimization for different application scenarios,and establishes a junior high school physical circuit diagram data set to verify the effectiveness of methods proposed.The specific research work of this article is as follows:(1)Propose a component recognition method for large-scale online services.Based on the deep learning algorithm,this paper chooses and compares several popular convolutional neural network.A multi-way grouping is proposed in this paper to solve the problem of complicated feature extraction network structure in the previous algorithm,which is time-consuming and not suitable for simple pattern recognition tasks.The structure's new feature extraction network ResNeXt28 verifies that the accuracy index of this method is superior to the traditional method in the circuit recognition task by comparison,which greatly improves the accuracy of the component recognition algorithm.(2)Design a lightweight circuit diagram recognition method for mobile terminal.In order to solve the problem that the image recognition algorithm model parameters are too large to be directly deployed on mobile terminals,this paper draws on a variety of network light-weighted strategies based on the general deep learning recognition algorithm,and randomly splits the feature channels in the neural network with reorganization,based on which a lightweight circuit recognition network YSNet is proposed.The results of experiment show that,compared with popular image recognition algorithms,YSNet not only greatly reduces parameters of the algorithm model,but also ensures the recognition accuracy,which means it has obvious advantages in deployment on mobile terminals.(3)Establish a physical circuit data set for APS.In order to verify the two circuit recognition methods above,this paper collected multiple exercise circuit diagrams from junior high school textbooks and authoritative examination papers,manually marks the location of circuit symbols and label,and then appropriately expands the data to establish a data set named Pscs-5500 which contains 5500 physical circuit diagrams.The experimental results show that the accuracy of diagram recognition has been greatly improved by the proposed methods.Besides,the model space occupation for both scenarios was reduced,which can be widely used in circuit diagram recognition for APS to improve the intelligence of online tutoring.