Classification And Recognition Of Dongba Characters Based On Convolutional Neural Network

Character recognition is a technology that combines pattern recognition,image processing and word processing.It is an important branch of artificial intelligence and pattern recognition.Through many years of research and exploration,character recognition has achieved satisfactory results in the recognition of English,German,Latin and other Western characters and Chinese characters.As the only hieroglyph still in use in the world,Dongba characters have unique historical value and rich cultural connotations.However,the classification and recognition of Dongba characters started late and the research is not deep enough.In the field of character recognition,the traditional method of feature extraction and classifier is often used to study.This method relies on specific algorithm to extract character features.The feature extraction algorithm is too pertinent and has poor generalization.The emergence of deep convolution neural network simplifies feature extraction,and the network can automatically learn effective features from data.Therefore,great progress has been made in the field of character recognition.Combining image processing and data enhancement,this paper proposes a method of Dongba character classification and recognition based on convolution neural network.The work done in this thesis is as follows:(1)Aiming at the problem of fewer Dongba character data sets,the original data sets are constructed by combining Dongba ancient books and Dongba character input method;then the Dongba character images are divided into single Dongba characters by pre-processing operations such as binarization,character segmentation and normalization;finally,the original data sets are processed by using data enhancement methods,such as diffraction transformation and noise jitter.The original data set was expanded from 956 to 30592.(2)For the constructed Dongba character dataset,four different convolutional neural networks(ResNet-18,VGGNet,AlexNet,LeNet)are used to classify Dongba characters.The experiment shows that the ResNet-18 network achieves better classification accuracy on Dongba character classification problem,and further improves the classification accuracy by adding pre-training network and changing optimizer,and finally achieves 92.3%classification accuracy.(3)In the process of studying Dongba character semantic recognition,because of the small number of Dongba character samples and the large number of semantic categories,we adopt the method of one-shot learning,and design siamese networks to compare the transformed shape representation and word vectors,and collaboratively carry out Dongba character semantic recognition.The experimental results show that the network achieves 85.6%recognition accuracy on the data set constructed in this thesis.To sum up,this thesis constructs Dongba character data set artificially,applies convolutional neural network to Dongba character classification and recognition research,and achieves good classification and recognition accuracy.It plays a positive role in the protection and inheritance of Dongba culture,and also provides reference for other minority character recognition issues.