Research On Detection And Recognition Of Chinese Characters In Any Direction In Natural Scene Pictures

In today's society,mobile phone photography has become the main way for people to record their lives,and a picture often contains rich text information in addition to object information.If the text information in the pictures taken by the mobile phone can be extracted according to people's intentions,it will be very helpful for people to obtain more effective information content.But there are many difficulties to be solved here.From the perspective of detection,the direction of the scene text is arbitrary,not a fixed horizontal direction,and their sizes are also different.In addition,the pixels of photos taken by mobile phones today often reach tens of millions.All these bring challenges to the accuracy of model training and the speed of inference.From a recognition perspective,unlike simple English text recognition,English has only 26 characters,and there are many types of Chinese characters.Secondly,the font types of Chinese characters in scene texts are different,such as all kinds of artistic characters.Coupled with the complex background of text information,these have brought great difficulties to the recognition of Chinese characters.The scene selected in this paper is a lot of storefront information on outdoor streets and text information on other outdoor buildings and street signs.From the perspective of practical application,there are two major directions that need to be balanced.One is to consider the size and speed of the model.We hope that the smaller the model,the better,and the faster the speed,the better,so that it can be deployed and applied on the mobile side.The second is to hope that the effect of detection and recognition can have a higher accuracy.From the perspective of the above two practical applications,the research contents completed in this paper are as follows:1.Lightweight detection model.The text detection model in this thesis is optimized on the basis of a model based on progressive scaling algorithm.The size of the model trained with the original model reaches more than 700 M,which is very unfavorable for the deployment of the model.Therefore,without reducing the accuracy of the model,the model is optimized,and its backbone network is replaced with a more lightweight network structure.The final model size is reduced by more than ten times,and the speed of calculation is also significantly improved.2.Improve the detection ability of dense text in any direction.The pixels of the pictures taken by mobile phones can only be trained by zooming and then detected and recognized.However,the zoomed pictures will inevitably make the original small text more difficult to detect.Although the fixed settings of the original model are helpful for the3.detection of dense text in any direction,but In the street scenes captured by the mobile phone,there are still some dense text in any direction after zooming,which is difficult to detect or detect the occurrence of sticking.Based on this,this paper optimizes its corresponding parts to improve the detection ability of the corresponding dense small text.4.The text recognition part enhances the model robustness by generating simulation data.In the text recognition part,due to the wide variety of Chinese text information on outdoor storefronts,outdoor buildings,and street signs,with different fonts,such as traditional and artistic characters,the background information of the text is also rich and diverse.If you only use the training set,you cannot cover all situations.Therefore,the strategy of generating simulation data is adopted to cover all possible text information as much as possible,thereby greatly improving the recognition effect.