Research And Implementation Of Calligraphy Recognition System Based On Deep Learning And Transfer Learning

As image digitization develops,researchers begin a further study in calligraphy recognition,which shows great value in the annotation of calligraphy image and study of famous calligraphy.However,recognizing calligraphy is difficult now,as there is no open dataset of enough calligraphy images and meanwhile the current calligraphy images are not so clear.At the same time,image recognition based on deep learning and transfer learning becomes a research hotspot in the field of computer vision recently,for its strong ability of generalization in small dataset and high recognition rate.To complete calligraphy recognition in this research,image recognition based on deep learning and transfer learning is combined with calligraphy images to analyze the requirements of calligraphy recognition in application,and a large number of calligraphy images of single Chinese character are collected for the extraction of high-level image feature in the convolutional neural network after transfer learning.First of all,the collected dataset of calligraphy image is preprocessed for later model training;Then,the model and weights trained in large dataset is transferred for the feature extraction of high-level calligraphy image and for the establishment of network model;Finally,sort the recognition probabilities of calligraphy calculated by the well-trained network model to get its corresponding recognition result,and display similar calligraphy image and the meaning of related Chinese character.The system will bring great convenience to the digital annotation of calligraphy images or the study of calligraphy works.The calligraphy recognition system is researched and implemented from two aspects:Firstly,deep neural network is combined with model-based transfer learning.Train the convolutional network on ImageNet with a large number of image datasets to get the model and its corresponding weights,which is used to initialize the weights of the same model in calligraphy recognition.Then train the network model with the initialized weights from the scratch under MXNet,the deep learning framework,on the dataset of calligraphy images until the network converges.At last,the well-trained model can not only recognize 6,760 Chinese characters and 5 calligraphy styles but also have higher recognition rate and practicability than related researches of calligraphy recognition under most metrics.Secondly,to construct the online system of calligraphy recognition,display and segmentation under the architecture of AntDesign+ VueJS + Spring +Mybatis,the model mentioned above is combined with the dataset of calligraphy images and Chinese character meaning.The system shows the operability of calligraphy recognition system based on deep learning and transfer learning intuitively and promotes the development of the field of calligraphy recognition.Meanwhile,calligraphy recognition based on deep learning and transfer learning is in the initial stage of development currently.Thus,the work done and its shortcoming are summarized,and the direction that can be further studied is prospected.