A Study On Object Recognition Of Questionnaire Images

Questionnaire surveys play an important role in excavating facts and developing effective strategies in various industries.In the questionnaire survey,the collected questionnaire is an important way to obtain relevant information.Although electronic questionnaires are widely used with the development of the Internet,in some research projects,paper questionnaires are still difficult to be replaced by electronic questionnaires.After recycling the paper questionnaires,the input of the answers mainly relies on manual methods.If the number of returned questionnaires is huge,it will obviously consume a lot of manpower,and the accuracy of the input is difficult to guarantee.Therefore,it is urgent to realize the automatic identification of paper questionnaires.At present,the target detection algorithm has achieved end-to-end recognition,that is,it can directly detect the area and category of the target.The YOLOv3 algorithm takes into account the speed and accuracy of detecting targets.Therefore,this paper applies the YOLOv3 algorithm to the target detection of the questionnaire to realize the automatic identification of the questionnaire answers.The main work of this paper is:(1)Issue and collect questionnaires of the same template and different templates.The questionnaire is designed to be given in parentheses for filling in the answer section.The answer category contains four categories: A,B,C and D.The collected questionnaires were photographed and saved,and the annotation of the handwritten answers in the questionnaire images was completed.(2)The YOLOv3 network is trained using the same template questionnaire,but the recall rate is low when testing questionnaires with different templates.After adding different template questionnaires to the training set,the precision rate,recall rate and m AP all reach more than 90(3)Looking at the questionnaires with poor detection effect,it was found that the questionnaires were mainly inclined,horizontal,and inverted.Therefore,data enhancement was performed on the training set to improve the detection effect of the YOLOv3 model on the inclined,horizontal and inverted questionnaires.The m AP(mean Average Precision)reaches 99.89%.