Design And Implementation Of The Objective Question Recognition System Of Answer Sheet Based On Machine Learning

At present,examination,as a method of checking the learning achievements in stages,is an indispensable part of the school teaching process.However,the review of objective questions is monotonous and repetitive,which wastes a lot of teachers’ time and energy,and is not conducive to analyze the result of the examination.This paper intends to develop a machine learning-based objective question recognition system for answer sheets,which can be compatible with multiple types of answer sheets.It has the advantages of convenient use and wide practicability,and is easy to promote in various schools,reducing the burden of teacher marking.Improve teaching quality and promote the informatization process of examination teaching.The main contributions of this thesis include the following aspects:Firstly,aiming at the widely existing answer sheets without positioning points and synchronization heads,we design a set of filling point positioning algorithms that do not rely on positioning points and synchronization heads.The algorithm is mainly based on convolutional neural networks,through the network finds out the position of the numerical question number on the answer sheet,and then accurately locates the position of the filling point according to the relative position relationship between the question number and the filling point.Secondly,in order to be compatible with various irregular fillings of candidates,we design a set of filling information recognition algorithm based on the comprehensive characteristics of filling points,which comprehensively utilizes the overall filling consistency,filling area and filling gray scale,etc.Using an improved K-means algorithm to identify filling information.Finally,we design and implement a set of answer sheet objective question recognition system based on machine learning.The system can easily identify answer sheets in batches,and can conveniently manage and query the test scores of candidates and classes.On the 100,000 real answer sheet image test data set,the accuracy of this system’s scoring has reached more than 99%,which shows the effectiveness of our algorithm and system.