Research And Applications Of Text Detection And Recognition In Natural Scene Images

Humans obtain information in various ways,and about 90% of the information comes from human visual information.In visual information,the text information that aggregates human crystals is often more informative than other information.Therefore,it is particularly important and meaningful to read and understand the textual information in the image.Traditional OCR technology has become mature in print recognition,but reading text information in natural scenes still faces great challenges.In natural scenes,unfavorable factors such as weak light conditions,poor weather conditions,and distorted text will affect our effective access to text information in natural scenes.In recent years,with the rapid development of computer technology and hardware,the technology of image detection and recognition in natural scenes based on deep learning has been greatly improved compared to traditional OCR technology.In the field of text recognition,the end-to-end CRNN text recognition model has always dominated,but the loss function CTC of this model is complicated and computationally intensive,and it cannot solve the problem of two-dimensional image recognition.Although it can be solved by the Attention mechanism,but the resulting text alignment problems reduce the accuracy of recognition.By introducing the ACE loss function,this paper can effectively solve the complex problem of loss calculation.The loss can also be extended to the recognition of two-dimensional images through the method of supervising the number of characters.The main innovations of this article include:1.Propose a text recognition model based on ACE loss.This model uses Dense Net as the basic network to extract text signs and send the extracted features to a bidirectional GRU to further extract the context features of the text.Finally,the loss calculation is performed by the ACE loss function.The model shows a performance that rivals CRNN in the recognition of one-dimensional images,and can be extended to two-dimensional images for text recognition.2.An end-to-end model for text detection and recognition in natural scenes is proposed.This model is based on EAST's text detection method.It obtains text regions through FCN and sends the obtained text regions to our ACE loss-based recognizer to uniform training.The model achieves good results in text detection and recognition.In the model evaluation,the F-score of the model is greater than 50%.3.An end-to-end natural scene text detection and recognition system was developed using Py Qt5 development tools.The system includes image information acquisition module,image detection and recognition module,image processing module,and display of processing results.It can not only complete the basic detection and recognition of the image,but also select the different detection or recognition model to compare the effect of detection or recognition.