Dongba Character Recognition Based On Multi-Scale Feature Fusion

In Lijiang City,Yunnan Province,China,there lives an ancient ethnic group,called Naxi nationality.With the wisdom and courage of Naxi people,they have nurtured the Dongba culture with its unique flavor,which has had a profound influence on the development of ancient Chinese culture.The Dongba hieroglyphs,one of the representatives of Dongba culture,are mainly used by the Dongba of Naxi people for many generations,so they are also called Dongba characters.the number of people who can read Dongba is decreasing,and the Dongba characters are facing the situation that no one can read them,Therefore,the digital construction of them is imminent.As a key part of the digitalization of Dongba characters,efficient and accurate identification of Dongba script is of great historical and cultural value.The paper proposes a Dongba character recognition model based on multi-scale feature fusion.Compared with other recognition models,the advantage of the model is that it fully considers the influence of a large number of similar characters in Dongba characters and uncontrollable factors on recognition in natural scenes,so as to improve resnet34 model,The model has a high recognition accuracy and strong robustness.The main work of the paper includes the following aspects:(1)Organize and construct Dongba data sets related to the Dongba characters.Based on the related literature,there are no official Dongba-related datasets that has been made publicly.The thesis uses the ancient Dongba book "Genesis" and the Dongba text input method(electronic Dongba)to collect and construct the datasets.Finally,three data sets were obtained,including: ordinary character dataset DB1424,similar character dataset DBS20,and natural scene character dataset CDB1.The DBS20 uses a trained Siamese network to detect the similarity of similar texts,which can make the data set more credible.(2)Propose a Dongba character recognition model based on multi-scale feature fusion.There are many similar characters in Dongba characters,such as similar characters in Chinese characters.If only use the existing recognition model for recognition,the existence of similar characters will seriously affect the recognition efficiency.In addition,as Dongba characters have a strong pictorial character with obvious shape features,the Histogram of Oriented Gradients(HOG)feature,which has more structural information,is considered to be incorporated into the recognition model.In view of the above situation,The paper proposes a recognition model based on multi-scale feature fusion for Dongba script.The recognition model uses Res Net34 as the base network,and fuses the shallow structural features and deep semantic features output from the network with the extracted HOG features to form the multi-scale fused features as the input of the classifier and output the results.The recognition model uses multi-scale feature fusion can learn more detailed features and has stronger feature learning capability.(3)Design and develop a mobile photo-recognition APP for the Dongba script,or "Dongba Tong" for short.Since similar OCR systems are available for Chinese,English and other languages,there is no such OCR system for ancient scripts,so the design and development of "Dongba Tong" is highly applicable and necessary.Based on the above-mentioned Dongba character recognition model and method,the "Dongba Tong" designed and developed by the paper can upload pictures of unrecognized Dongba characters to the back-end server.After the server receives the Dongba characters pictures from the mobile terminal,the recognition model is adopted for recognition processing,and then returns the recognition result to the mobile terminal and displays it on the user interface for user reference.