Research On The Detection And Recognition Algorithm Of Dongba Character Based On Deep Learning

China is a multi-ethnic country with many nationalities,languages and characters.Various kinds of characters created the diversity of Chinese culture.Nowadays,many languages and characters are in danger of disappearing.It is of great historical significance to protect the diversity of languages and characters.Dongba is a kind of characters used by the Naxi minority.Nowadays,fewer and fewer Naxi people can recognize and write this kind of characters,which will finally lead to the loss of Dongba.In this paper,a model based on deep learning,which integrates detection,segmentation and recognition,was established to detect and recognize the handwritten Dongba images in scenes.First of all,based on the characteristics of Dongba characters,this paper collected and wrote Dongba characters and constructed a valid dataset for character classification,and collected 500 classes with a total of 3000 samples.In addition,this paper collected domestic and foreign open-source text detection datasets for pre-training,meanwhile collected and annotated the pictures of scriptures and sculptures as the fine-tune dataset with a total of 30 pictures containing 221 words.Before training model and model reasoning,the image was preprocessed in this paper.The processing flow is stretching and scaling the image to a uniform size,smoothing the image by median filter,and converted images to binary ones by using adaptive threshold method.Secondly,this paper selected EAST model for feature extraction and merging after comparing many papers.In addition,in the way of numerical experiment analysis,this paper compared,analyzed and verified the methods of text recognition,such as template matching,support vector machine,deep learning,etc.Finally,it selects deep learning for recognition since the highest efficiency.The experimental results show that,the recognition accuracy of hand-written Dongba dataset is up to 100%,and the recognition accuracy of scene Dongba characters is 93.5%?Finally,after model training and integration,a Dongba detection and recognition model with the best parameters is obtained.