Research On Chinese Scene Text Recognition Based On Deep Learning

Text recognition is a general image understanding technology,which is of great significance to the research of information retrieval,automatic driving and other applications.Thus,scene text recognition has gradually become a hot issue in computer vision research,while Chinese scene text recognition is one of the most important and challenging tasks in image recognition.Compared with English characters,Chinese Characters have complicated structures and a large variety,which is difficult to recognize.With the great revival of neural networks,scene text detection and recognition tasks have been greatly promoted.In recent years,many algorithms for scene text detection and recognition based on deep learning have emerged,but most of them are oriented to English datasets,while researches on Chinese scene images are relatively lacking.This paper studies the representative algorithms in the field of text detection and recognition.In view of the research and practical requirements of Chinese scene text recognition,this paper mainly carries out the follwing research work: first,we choose excellent algorithms of scene text detection and recognition in English datasets,with which do a large scale experiments and comparative analysis on Chinese scene datasets,and then summarize the challenge of Chinese scene text recognition;Secondly,for the problem that multi-directional text lines in natural scene are not easy to recognize,we cut the text lines from the original image by perspective transformation according to quadrilateral coordinate points;And then,according to the problems of current text detection and recognition algorithms in Chinese scene images,we design a Chinese scene text recognition algorithm based on character detection.At last,we prospect the future work in the field of scene Chinese character recognition.At first,we do large-scale experiments and performance analysis of the popular scene text detection and recognition algorithms on common English and Chinese scene datasets.For the text detection part,EAST and Text Boxes++ are used to experiment on two English datasets and four Chinese datasets,and cross-data and cross-language tests were conducted to explore the influence of texts' language on scene text detection through experiments.At the same time,in the text recognition part,we choose three text recognition algorithms for research which are sliding CNN,CRNN,ASTER,and we proposed an improved algorithm based on Sliding CNN,which is named Slice CNN.Among them,in order to compare the effect ofdifferent feature extraction networks on text recognition,we use three basic neural network structures in the feature extraction of Sliding CNN,Slice CNN and CRNN,which are VGG,Res Net,Dense Net.These ten text recognition algorithms were used to do experiments on three English datasets and four Chinese scene datasets respectively.Secondly,to solve the problem that multi-directional text is difficult to be recognized,a text line rectification method based on perspective transformation of original image is proposed.In the original image,the perspective transformation is used for rectification,and then the text line is cut out from the transformted original image,which can increase the recognition accuracy by about 12%.Compared with the previous method,which is according to the minimum external positive rectangle of the quadrilateral,and then the recognition is carried out after correction by using the spatial transformation network,it avoids extra training and key point prediction,simplifies the learning of the model and saves the training time.At last,in order to avoid the problem of lacking large-scale training samples,we design a scene text recognition algorithm based on character detection,which transforms the sequence recognition into the problem of instance segmentation and image classification.Considering that there are many distortions caused by perspective in the scene image,the perspective transformation is used on the character area after the character detection.The experimental results show that the algorithm is effective.In summary,this paper analyzes the limitations and defects of the existing text detection and recognition methods on Chinese scene text datasets,and puts forward corresponding solutions to Chinese scene text recognition,which has certain reference value for the research of scene Chinese text recognition.