Research On Application Of Calligraphy Character Recognition Based On Deep Learning

Nowadays,text detection and recognition in natural scenes has become a hot field in computer vision.The goal of character detection is to get the position of the Chinese line in the image through training,while character recognition is to extract the characters from the text line.Compared with character recognition in general scenes,the detection and recognition of irregular ancient Chinese calligraphy characters is more challenging because of its long history and lack of proper protection;and different from scene character recognition,there is no dataset about irregular calligraphy font at present;moreover,the ancient calligraphy fonts cover rich contents,including seal script,official script,regular script,running script,cursive script and other forms,which are quite different from modern Chinese characters;at the same time,the typesetting of irregular calligraphy font is Vertical,this is also different from the horizontal text line which accounts for a huge part of scene character recognition.To solve the above problems,this paper establishes an irregular character image dataset,improves the existing character detection model,and uses two methods to realize character recognition,and finally built an irregular character recognition system.First,this paper establishes an irregular text image dataset,which includes 8500 images of irregular calligraphy text in natural scenes and labels the position and content of the text.Most of the fonts in the dataset are regular script and running script,and a small number of cursive script is also included.Secondly,aiming at the problem that the existing text detection model cannot detect the vertical text column well,this paper improves the existing two models to solve the problem of text detection.Then,this paper uses two models of CRNN+CTC and attention mechanism in the stage of character recognition and analyzes the experimental results.Finally,the irregular character recognition system is designed and implemented.This system is based on the highest accuracy model in the previous two stages of text detection and recognition.Users can detect and recognize irregular text and natural scene text with simple operations.