Deep Learning Technologies In Paper Score Detection And Recognition

In school,score collecting and statistical measurements after tests is a necessary part of analyzing students’ performance and improving teaching quality.The automatic recognition of handwritten scores on test papers can effectively reduce the workload of typing scores on pieces of paper into the computer so as to improve work efficiency.This paper implements an end-to-end deep-learning based test paper score extraction application,designs methods for test paper handwritten score detection and recognition,and implements a ready-to-use test paper score automatic recognition Web application.From the design of the algorithm to the implementation of the application,the research results with a practical engineering product are given.The main work of the paper includes:1.Collected more than 8,000 test papers,and marked the score position and score of more than 700 of them.At the same time,collected 3 8,167 integer images covering the range of 0～100 and scores with 0.5 points,which provides sufficient data for deep learning networks;2.To solve the problem that handwritten scores are relatively small compared with the test paper,therefore are prone to be missed,a score column detection algorithm based on the combination of traditional image processing methods and deep learning methods is designed,which can effectively extract the score column in the test paper.And the running time of the detection algorithm is half that of the method of double deep learning network when the accuracy is basically the same;3.According to the image characteristics of the handwritten scores of the test paper,the network structure of CRNN has been modified,which can realize the automatic recognition of the handwritten scores of the test paper.Besides,a training paradigm is optimized for noninteger scores,which can make the CRNN network have a better recognition effect on noninteger scores.The experimental results show that the integers in the range of 0～100 and the decimals with the granularity of 0.5 can be well recognized,and it has good robustness to the noise such as falsification,horizontal straight lines and non-digital writing;4.A web application for automatic recognition of test paper handwritten scores is designed and implemented.It provides clear page information guidelines and a simple and elegant user interface.Functions such as image upload,score detection and recognition,and results download are provided.Users can use any device to log in to the website.