Research On Booklet Scanning Robot Equipment Control And Text Recognition And Error Correction

Booklet scanning and imaging is an important channel to realize digitization,and using intelligent and fully automatic booklet scanning robot is the best way to complete the digitization of books and journals efficiently.This research is about a booklet scanning robot equipment,which can automatically turn the pages of books and magazines,and realize the text recognition and text correction of scanned images based on deep learning technology.The booklet scanning robot equipment consists of an industrial camera,an automatic page turning device,a "V" shaped support and a book carrying base.Based on the original equipment,the automatic page turning device is optimized,based on the principle of vacuum adsorption,using the swing arm page turning device,with the "V" shaped support,to improve the reliability of automatic page turning.The control system hardware adopts Siemens S7-1200 PLC and V90 servo system,using servo motor to realize the up and down reciprocating motion of "V" shaped bracket and automatic page turning device in and out and turning motion.The vacuum pump and solenoid valve are used to control the single page adsorption,and after the page is turned,the "V" shaped bracket restrains the position,and the industrial camera perpendicular to the two inclined surfaces of the "V" shaped bracket takes pictures and scans to obtain the scanned pictures of the two adjacent pages of the booklet.The PLC and V90 servo system are configured on Siemens TIA software for hardware configuration and communication settings to complete the automation program development.According to the actual demand of booklet scanning,in addition to acquiring booklet pictures,text data is also acquired.In this paper,based on deep learning technology,a text detection model based on DBNet neural network and a text recognition model constructed by CRNN+CTC network are used,and optimization strategies are proposed.In terms of learning rate,Cosine learning rate strategy is introduced to make the model converge faster.The text detection model is optimized with the SE module to reduce the model size and improve the inference performance while maintaining the accuracy,and the final validation set detection accuracy is88.09%.The text recognition stage introduces data enhancement and pre-training model,which experimentally shows that it can effectively improve the model accuracy,and the recognition accuracy of the validation set is 94.91%.For the situation of morphological word errors after Chinese text recognition,this paper proposes a text error correction method based on MacBERT and word bar codes,which is mainly divided into three steps: error detection is a threshold screening by text recognition confidence,and the MacBERT model is responsible for contextual semantic recall to select candidate words based on suspected error words.The main idea of word barcode is to use Chinese character features for encoding,find out the character similarity according to the feature weight calculation formula,and finally synthesize the semantic recall model score and character similarity score for ranking to get the optimal result,and the final validation set error correction F1 value is 71.75%.The text recognition and text error correction models are inferentially encapsulated and the application is developed using Qt and Flask development framework.The upper computer application realizes the control of the lower computer of the booklet scanning robot through TCP communication,and the text recognition and text error correction application realizes the text recognition and text error correction of local pictures.After experimental verification,the application system meets the practical requirements,and the research content of the topic has certain engineering application value.