Multi-scale Feature Fusion Based Dongba Character Recognition

Abstract

Dongba characters are hieroglyphs still used by The Naxi nationality in Yunnan province, and have a long history which is called "living fossils" of characters. In view of the neglect of the influence of similar characters on the recognition accuracy and the absence of the construction and application of relevant models in the existing studies on Dongba characters, this paper proposes a multi-scale feature fusion based the character recognition model for Dongba characters. First of all, due to the lack of Dongba characters data set, two Dongba characters data sets (including DB1424, a common data set, and DBS20, a data set about similar characters) are constructed. Then, the design of this model mainly uses ResNet34 as the basic network model. While extracting the ResNet34 network features of the input image, the traditional feature extraction algorithm is used to extract the directional gradient histogram (HOG feature) of the image that can better reflect its structure and shape. The structural features of shallow layer and the semantic features of deep layer in the network are fused with HOG features to form a combined feature, and finally the Softmax classifier is used to output the results. The results of the experiments illustrate that the accuracy of the recognition model proposed in this paper on DB1424 and DBS20 are 89.712% and 94.729% respectively. Compared with other deep learning models, the model proposed in this paper has better learning ability, and the precision and recall rate are also higher.