Research On Handwritten English Character Recognition Technologies For Examination Paper Based On Deep Learning

In recent years,computer technologies have become increasingly mature,and more and more intelligent and digital products have been widely used,which has brought great convenience to human life.At present,with the high development of education field,the number of various examinations is increasing.In traditional scenarios,the answers of examinees are mostly stored in the form of examination answer sheets.However,this kind of storage has certain limitations,such as it is difficult to achieve permanent storage and unable to make further analysis of the examinees’ answers.Thanks to the continuous development of text recognition technology,the contents in examination answer sheets can be transcribed into text form that can be understood by computers,which solves the practical problems faced in this field.By collecting the datasets of examinees’ English answer sheets in real collage examination scenarios,and then using the relevant technologies in the field of text recognition,the contents in examination answer sheets can be recognized and transcribed.It provides the possibility for the permanent storage of examinees’ answer sheets and the automatic marking technology in the future,thus speeding up the intelligent process in the field of education.For the different writing situations that may occur,in this thesis,different methods for handwritten English character recognition are proposed,including word-based handwritten English character recognition and line-based handwritten English character recognition.The main contributions are as follows:(1)Word-based handwritten English character recognition.In order to ensure the accuracy of recognition,for examinees whose writing is more regular and the answers can be segmented into words,in this case,the word-based handwritten English character recognition can be conducted.For the purpose of allowing the model to additionally use the information of the importance of each channel in feature maps,the channel attention module is proposed in the visual feature encoder.At the same time,in order to fully combine the advantages of different decoding methods,in the visual feature decoder,a multi-mode visual feature decoding method is proposed,which combines the decoding method based on connectionist temporal classification and the decoding method based on attention mechanism.It can speed up the model convergence speed so as to improve the model recognition performance.(2)The proposed word-based handwritten English character recognition model,it is trained and tested on the corresponding word datasets.The comparison experiments with mainstream methods and the ablation experiments of each module in the model are carried out,and finally the recognition results of the model are visualized.These results show that the word-based handwritten English character recognition model can achieve satisfactory recognition results.(3)Line-based handwritten English character recognition.In some situations,the words in examination answer sheets cannot be segmented due to the irregular writing of examinees.In this case,the word-based handwritten English character recognition model is no longer the optimal choice.In order to make the model more universal,it is necessary to recognize handwritten English character with longer text length,that is line-based handwritten English character recognition.According to the characteristics of English line datasets,the line-based handwritten English character recognition model is obtained on the basis of the word-based handwritten English character recognition model.Different from English words,English lines include word-level information in addition to character-level information,and there is abundant semantic information in English lines.Therefore,a vision module that can realize multi-granularity visual feature decoding and a semantics module that can extract latent semantic information in English lines are proposed.Since visual information and semantic information can describe different aspects of images,for the purpose of making full use of the extracted visual information and semantic information,a fusion module that can fuse the information of these two modalities is proposed.Finally,the fused features are used for prediction.(4)The line-based handwritten English character recognition model,is trained and tested on the corresponding English line datasets.The comparison experiments with mainstream methods and the ablation experiments of each module in the model are carried out,and finally the recognition results of the model are visualized.These results show that the line-based handwritten English character recognition model can achieve satisfactory recognition results.