Research On Text Detection And Recognition In Natural Scenes Based On Deep Learning

Text detection in natural scenes refers to a technology for locating the bounding box of words or text lines in images.In recent years,as the society's demand for text detection in natural scenes has increased,it has promoted the rapid development of scene text detection.As a part of scene text detection,text detection has gradually become an indispensable technology in this development process.It has broad application prospects,such as automatic driving,security,traffic control systems,blind assistance systems and other fields.application.Due to the diversity of scene text and the complexity of the background of natural pictures,scene text and recognition is undoubtedly a challenging task.Based on deep learning technology,this access proposes solutions to the problems of scene text detection and recognition.The main work of this access is as follows:(1)Propose a scene text detection method based on automatically generating guided bounding box module.This method uses a single-stage text detection method to automatically generate a guided bounding box,which is used as the input of the second-stage text detection to further correct the boundary of the text bounding box and remove the wrong text bounding box.The automatically generated bounding box is more in line with the shape of the text,and the parameter adjustment of the predefined bounding box is omitted.Experimental results show that the scene text method proposed in this access effectively reduces the false detection rate and improves the accuracy of the detection results.(2)Propose a scene text recognition method based on a supervised text correction network.The method focuses on the correction of picture text,which can correct vertical text and irregular text to normal horizontal text.For vertical pictures,this article designs a vertical text correction module to correct vertical text into horizontal text.(3)For the corrected vertical pictures and other pictures,this access uses a novel supervised spatial transformation network to correct irregular characters.The corrected picture is input to the encoder based on the convolutional recurrent neural network for feature extraction,and finally the encoded features are converted to the final output through the sequence-to-sequence model decoder based on the attention mechanism.The scene text detection method based on the automatically generated guided bounding box module proposed in this access can improve the detection accuracy in conventional natural scene text detection.The text correction network can effectively correct vertical text and irregular text into normal horizontal text,and improve The recognition rate of the recognition results.Therefore,the research in this acccess plays an important role in improving the efficiency of text detection and recognition in natural scenes,and better serving the needs of text detection in natural scenes.Unfortunately,this access does not significantly improve the detection accuracy of multiple cross-language scene text detection and very complex deformed text,and further work is needed.