Research And Implementation Of Invisible Code Text Recognition For Traceability Of Cigarettes Based On Deep Learning

OCR is one of the popular directions in the field of computer vision,and its products and applications can greatly facilitate our daily work and life.Traditional OCR technology has been well developed for text recognition of documents,but in natural scenes,due to complex backgrounds and diverse text,the image text recognition for general scene is still a challenging task which requires continuous investigation and improvement.For now,scene text recognition models are more often landed on a certain situation to achieve a better result.Therefore,this thesis takes traceability of cigarette based on Invisible Code as the application background,takes scene text recognition on mobile devices as the main research content.Therefore,this thesis takes traceability of cigarette based on Invisible Code as the application background,and takes scene text recognition with mobile devices as the main research content,optimizes the model structure and training strategies for deep learning models applied on mobile devices.In addition,this thesis designs the mobile application and the Integrated Platform which supports it.The main work are as follows.1.Organizing and labeling the Invisible Code dataset to solve the problem that general scene text dataset is very poor in application.The Invisible Code is a kind of text that can only be seen under special lighting,and it’s shape is different from the usual text in life.2.Analyzing the evaluation criteria in the general scene,defining the evaluation metrics which are suitable for the invisible codes of cigarettes.By comparing several models in both phases of OCR according to the new evaluation metrics,By comparing several models in both phases of OCR according to the new evaluation metrics,we select the DBNet and CRNN models as base models because of their comprehensive performance.3.Constructing a invisible code recognition model by optimizing the model structure and training strategies of base model which is suitable for mobile devices.With guaranteed recognition accuracy and compared to base model,the model size is reduced to about 1/7,and the runtime of mobile device is reduced to about 1/3.4.Proposing an Invisible Code recognition correction processing method.This method uses Bi LSTM model and heuristic correction repair strategy to check and restore the information of Invisible Code,and it can it can improve the accuracy when Invisible Code parsing.5.Designing and implementing the "Invisible Code Recognition APP" and integrated platform for Invisible Code management.Applying the deep learning technology to the traceability process of cigarettes.