Research On Algorithm Of Extracting Wrong Numbers From Examination Papers Based On Deep Learning

In the past ten years,with the continuous improvement of national quality,examinations have become an important standard for testing teaching results.The continuous development of computer science and technology has changed the traditional manual scoring method,and network scoring has become an important part of the examination process.During the scoring process,the teacher needs to judge whether the subjective question is right or wrong,and input the score of the question,and finally the score will be counted by the computer.If the computer can directly detect the error number and give a score,the workload of the teacher can be greatly reduced.This thesis uses the deep learning method to detect the wrong number in the test paper,and proposes an improvement to the algorithm for the test paper data set and obtains a high-accuracy detection result.This paper first studies the traditional target detection algorithm.The traditional detection method is to manually design the feature extractor,which has poor detection effect and slow speed.Therefore,this paper chooses the representative Faster R-CNN algorithm among the two-stage algorithms as the basic framework.And improve it.Use Mobile Net V2 network and Res Net50+FPN network two different backbone networks to train the data set to compare the quality of the two networks.Changing the pooling layer from Ro I Pooling to Ro I Align avoids the errors caused by the two quantizations and improves the detection accuracy.Since there is currently no public test paper data set,this article collects a total of 1200 test paper data sets,and labels them and processes them into PASCAL VOC format data sets.Finally,the Res Net50+FPN network with good detection effect is used as the feature extraction network for feature extraction,the K-means clustering algorithm is used to determine the anchor size ratio suitable for the test paper data set,and the bilinear difference method is used to obtain different sizes on the feature map.The feature value at the floating point coordinate position in the candidate area,and finally the multi-classification function of the fully connected layer is changed to the two-classification function.The experimental part compares the experimental results of Faster R-CNN algorithm using Mobile Net V2 and Res Net50+FPN two different feature extraction networks on the PASCAL VOC2012 data set and test paper data set.The average accuracy of the mean value on the PASCAL VOC2012 data set has reached respectively.0.666 and 0.737,the average accuracy of the mean value on the test paper data set reached 0.970 and 0.990 respectively,which shows that the Res Net50+FPN network has a good effect and has reached a very high accuracy rate.The experimental results show that the method of extracting wrong numbers from test papers based on deep learning proposed in this thesis has a high accuracy rate,and has certain research prospects in the future research direction of network scoring.