Research And Deployment Optimization Of Text Recognition In Scenes With Large Angled Curves

In recent years,text recognition technology in natural scenes has received increasing attention.Scene text images are complex and varied,with many interferences,and the text itself has various shapes,especially for text with large curved angles,where the rotation angle leads to similar shapes between different characters,making it difficult to distinguish.Even the same character can be difficult to converge due to significant differences,which poses a great challenge to accurate text recognition.To address this problem and meet the industrial deployment needs of the algorithm,this thesis mainly focuses on the research of text recognition algorithm design and the construction of efficient computing modes.Firstly,this thesis designs a multi-angle fusion character recognition algorithm,which divides the angle domain into multiple sub-domains and uses sub-networks for specialized training.Additionally,an adaptive angle perception module is designed to unify the sub-networks.Experiments show that the proposed algorithm improves the average accuracy rate by 1.9% compared to cutting-edge algorithms on multiple public datasets.Since there are few scene text datasets that contain character annotations,it is challenging to drive the development of character recognition algorithms.Therefore,this thesis compensates for this deficiency by annotating and correcting existing public datasets and developing a framework for generating character data automatically.Furthermore,starting from the industrial application of steel coil spray coding recognition,this thesis constructs a cloud-edge collaborative computing mode based on a voting mechanism and optimizes this framework from multiple aspects such as transmission schemes and task granularity parallelization.Finally,the proposed character recognition algorithm is deployed in the steel coil spray coding recognition application in the industrial cold rolling workshop.Therefore,the multi-angle fusion network designed in this thesis can effectively cope with the large span of rotation angles in scene text and has certain robustness on multiple public datasets.The proposed cloud-edge collaborative computing framework can effectively meet the low-latency and high-precision requirements of industrial scenarios.